Last week’s post left polation as a definition without much to hold onto: the discipline of smoothing an incomplete picture into a continuous whole, early enough that development can still be shaped by what you are learning. A definition is a start, but is not yet usable in a room full of people who have to decide what an asset is for.

Take sildenafil, briefly, since it is the case everyone already half-knows. An angina programme that was quietly, provisionally, also an erectile dysfunction programme, then a pulmonary hypertension programme, without anyone changing the molecule. The interesting question is not “what happened” - luck, in the press-release version. It is what kind of decision had to be made, twice, for that to happen. Someone had to decide, before the confirmatory trials had chosen for them, that the disease the molecule was being developed for might not be the only disease it was for.

That is a specific move. I argue for it a lot. Call it indication polation: what disease is this even for, asked while the answer is still open (the answer is mostly ‘still open’). It is not the only shape the discipline takes. The reason to separate the shapes is not my taxonomic obsession. A team can be superb at one kind - sildenafil’s programme was (Pfizer was good at this, a while back…) - and never attempt another. An organisation that only ever tries the kind it is already good at will mistake a narrow competence for a general one.

Patient polation asks which one physician and one patient the median is hiding - the Target Patient Profile: the obesity patient who was never going to present as a “non-compliant diabetic,” visible only because someone let the trial population be a question rather than an assumption.

Category polation goes further still. Not filling a gap inside an existing market, but redrawing where the market’s edges are - treating a diagnostic boundary or a disease definition as provisional rather than given.

Endpoint polation changes which measurement counts as knowing what you have. A biomarker or a response rate stands in for a disease category, so that the trial you can actually run this year tells you something the disease name alone would not. I’ve railed for a long time about 6MWTs and ADAS-COGs - they were no more valid then than they are now.

Language polation is the words that later become the label, value story and physician shorthand, decided early enough that they still have consequences rather than merely describing ones already locked in. (You know my other blog, Pharmaceutical Positioning, is all about this…)

Payer polation asks what would have to be true for this to be worth paying for, before you lock a trial design that cannot generate that evidence. Clearing a technical hurdle is the least the study can do.

Modality polation keeps the mechanism constant while changing the object around it: a pen versus a pill, weekly versus daily, diagnostic-required versus clinical standard practice.

That is seven kinds of polation, based on one verb. Indication, patient and category answer what this is for. Endpoint and payer answer what would have to be true. Language and modality answer what form the product takes while you are still allowed to change it. They are relatives, not a neat McK grid. We should use the split only so far as it stops a team from congratulating itself for the one kind it already does, or is used to doing.

The failure modes matter as much as the kinds, because molecules get polished by habit whether anyone intends it or not.

Default polation searches for analogy, fundability, last year’s TPP in the franchise - the product that requires no one to argue for it.

Overpolation is the opposite failure with the same cause: a beautiful story the data cannot yet carry, polished because the story is more fundable (or the competition seems to know something you don’t) than the honest, unfinished version.

Underpolation treats Phase I as an administrative prelude (it’s a safety check…) and postpones the product until the readout “tells you,” as though a p-value were a decision rather than an input to one.

Late polation is usually just repair - Commercial asked to invent meaning after the programme has already chosen, dressed up as positioning.

And depolation, the rarest of the five, strips a false story while the programme still looks healthy: killing a narrative that already has a champion and a colour-coded slide, before the organisation is embarrassed into it. Most organisations can add a scenario. Almost none can subtract one that already has a sponsor.

None of this (I hope) is abstract if you are willing to use it plainly. Put a possibility tree on an asset instead of a Target Product Profile: three to five live scenarios, one of them currently winning by default - by habit, not by evidence - named out loud so the room can see what it is already polishing toward. Fund one falsifying experiment per leading scenario, cheap enough that killing the story is rational rather than career-limiting. If every study on the books is designed to underpin the favourite, nobody is learning anything; they are rehearsing what they already believed. And keep the language provisional for longer than feels comfortable. The words for a molecule, once they have been through brand council, do not get revised because a biomarker is equivocal.

Case studies are coming - the argument does not hold up as a list of handy definitions, and the next piece (next week) is the products, the examples, not the theory. But I thought that the taxonomy is worth having in hand first, because otherwise every case reads as its own miracle. It is not a miracle once it has a name.

If you cannot name the default product of your own asset right now, you are not waiting for data. You are being polished by the organisation, and the organisation is not going to tell you.

The old question is still the useful one. If you gave the same asset to two teams, which set of products would you rather own when the picture finally fills in? The team that waited will own the median product. The team that polished early may own a faster one, or a second-but-better one, or one the first team never thought to look for.

You both inherit the same sparse early signals: a mechanism that looks interesting, a first-in-human curve that is more suggestive than conclusive, a patient need that is still being argued over in conference corridors. One team waits. They treat Phase I as safety and pharmacology, lock a single Target Product Profile because governance wants a North Star, and promise themselves they will “position” the asset once the pivotal data arrive. The other team starts earlier. Not by pretending the picture is complete, and not by stretching a few uncertain points into a peak-sales hallucination, but by smoothing those sparse signals toward a coherent sense of what this product might become.

That second team is polating.

Interpolate and extrapolate have been covering adjacent territory for years. To interpolate is to estimate the missing value inside the data you already have. To extrapolate is to project beyond those points. Both matter. Neither names the discipline that matters most in early development: shaping an incomplete picture into a usable whole while optionality still exists. The root is more useful than the prefixes. Polire: to polish. English inherited the directed verbs. It somehow misplaced the core action.

So: to polate is to smooth an incomplete picture into a continuous whole, early enough that development can still be bent by what you are learning. (I will claim credit - I can’t honestly find a reference or a useage for the word anywhere…)

That is not a branding exercise. It is a learning discipline.

The molecule is not the product. Polation is how you learn what the product is.

Asymmetric learning, as I have been using the phrase, is simple enough: learn faster and more creatively than the other team about what the molecule can do, what the market might want, and what might be approvable. The gap between a compound leaving the lab and a medicine reaching market is filled with learning. Most organisations do that learning symmetrically. They read the same reports, attend the same advisory boards, pursue the same endpoints, and then act surprised when the resulting labels look interchangeable.

Polation is the positioning-specific form of that asymmetry. It treats identity as something to be discovered while genuine flexibility remains, not something to be announced after the major choices have already been spent.

That is why it belongs in early development, next to optionality, rather than at the end of the process next to launch readiness. We have spent a fair amount of time arguing that early development should be a high-resolution option-generation engine: multiple paths in parallel, cheap kills, planned serendipity, Target Opportunity Profiles rather than a single TPP married in haste. That argument is usually made in the language of indications, biomarkers, and trial design. It applies equally to the question of what the product is for.

You polate when you refuse to treat first human data as an administrative prelude to the “real” study. You ask what the proximal effect could mean in a patient’s world while the dataset is still embarrassingly small. You generate several genuinely different hypotheses about category, use, and value while the programme can still be shaped toward them. You treat unmet need as something that can be framed with more intelligence than the current category structure allows. That is directed optionality applied to meaning, not only to protocol.

Organise the uncertainty. Do not wait for the evidence to commit on your behalf.

One of the quieter enemies of pharmaceutical innovation is the belief that a decision becomes more scientific if it is delayed until the aggregate picture is tidy. It feels rigorous. Often it is simply refusal in a lab coat. The median is not the message; the choice that matters still happens with one physician and one patient. Positioning has its own version of the same mistake: waiting until the programme has largely declared itself, then asking for a differentiated story as if differentiation were a communications exercise rather than a developmental one.

Polation does not eliminate uncertainty. It organises it. The point is not to force premature closure. The point is to ensure that indication choices, endpoint design, population framing, regulatory language, and value proposition begin to move in the same direction early enough to matter. Later evidence should be filling gaps in a shape that has already been prepared, not rescuing a programme whose meaning was left to chance, analogy, or whoever wrote the last governance slide.

That is the difference between chaotic learning and organised learning. We have only ever learned by doing. The question is whether the doing is pointed.

A rigid TPP fixed too early creates one kind of mistake: the illusion of certainty. A team then defends that certainty long after it has ceased to be useful, and everything downstream becomes more symmetrical - same endpoints, same populations, same commoditised claim. But waiting too long creates the opposite mistake: shapelessness. A Possibility Tree held deliberately through early evidence generation is the more useful posture. Three to five live scenarios, maintained long enough to be tested, is simply polation with a spreadsheet attached. You are not pretending the future is known. You are refusing to let the unknown remain formless.

Give the same early asset to two teams and the divergence is rarely raw talent. It is whether one team uses the foggy morning of development - biology still partly art, data sparse and noisy - to learn what the product could be. The other team uses it as a waiting room.

A useful example is sildenafil. The familiar story gets told as serendipity, which is true as far as it goes. But serendipity alone is not enough. Plenty of programmes throw off unexpected signals. The advantage comes from recognising that an early, incomplete picture may imply a different product identity than the one the original plan assumed, and then developing that implication rather than treating it as noise. That is the essence of polation: not prediction, but the disciplined shaping of possibility while the asset can still become something else.

Most late-stage positioning is repair. The programme has already chosen its endpoints, narrowed its population, habituated its language, and locked its comparator. Positioning is then asked to make the result feel inevitable. Sometimes a gifted team can still do it. More often the story is bolted on, and everyone can feel the bolts.

The deeper work happens in the weave. You polate when trial design is still a positioning instrument, not merely a regulatory instrument. You polate when “what might be approvable” and “what the market might want” are allowed to interrogate each other before the protocol is frozen. You polate when the first signals are treated as material to be worked with, not stains to be explained away later.

This is also why the verb is useful inside the company. The next time a team argues whether a framing is “supported by the data” or “too speculative”, the better question may be: are we interpolating, extrapolating, or polating? Are we filling known gaps, projecting beyond them, or shaping the incomplete whole so the product’s place begins to feel prepared?

Those are three different learning modes. They produce three different assets, even when the starting molecule is identical.

The uncomfortable truth is that molecules are already being polated whether anyone says so or not. Early assets are always being shaped: by default, by habit, by the first indication that looks fundable, by the analogue sitting in the competitive slide, by the language that happens to be lying around in the organisation. The polishing is happening anyway. The only real choice is whether it happens deliberately.

Asymmetric learning has always been a claim about velocity and direction, not about superior crystal balls. No one predicts perfectly. The winners learn faster, cheaper, and with more upside when the data cooperate. Polation is the name for how that advantage reaches the question most teams postpone for too long: what, exactly, is the thing we are making?

If the other team waits for Phase III to tell them what they have, they will learn what every other competent team can learn from the same readout. If you polish early - if you work with the gaps rather than fearing them - you may launch faster, or second but better, or in a place the first team never thought to look.

The verb is new. I coined polate for this. The practice is not. We already know how to interpolate. We already know how to extrapolate. The missing discipline is the one that happens while the picture is still incomplete, and while the molecule can still become more than one product.

The old question remains the useful one: if you gave the same asset to two teams, which one would you rather be on?

The thesis held. My explanation for it didn’t.

In July 2021 I described a molecule team that could stress-test twenty or thirty launch scenarios for a single asset in an afternoon - rapid progressors instead of all-comers, three different label strategies, three different price points - against the alternative: commission a market research exercise built around one Target Product Profile and wait three months for the binder. My claim was narrow: a good answer delivered in seconds beats a detailed answer delivered after the strategic window has closed, because the slow path over-weights whichever hypothesis happened to go in first.

In the original July 2021 piece, I argued that early-phase work is exploration, not prediction. A team that can give a good-enough answer in seconds to questions like “what if we targeted rapid progressors instead of all-comers?” can stress-test twenty or thirty real launch options in an afternoon. The industry default - commissioning a three-month research exercise against a single Target Product Profile - locks in the first hypothesis, arrives after the strategic window has closed, and produces the same obvious answers everyone else gets. Fast, experience-based answers create asymmetric learning; slow, detailed ones do not.

Five years on, that claim has held up better than most things I’ve written on this blog. What I got wrong was the explanation for why it would keep holding.

The obvious change is technology, so I’ll deal with it briefly. Agentic systems can now interrogate a decade of launch analogues, patent expiries, payer behaviour and competitive pipeline activity in the time it takes to ask the question properly. Scenario modelling that used to consume a fortnight happens inside the meeting where the scenario gets asked. None of this was available in any practical sense in 2021; all of it now is, to more or less anyone.

The reading I’d have given you in 2022 is that the constraint that used to force teams to evaluate two or three options instead of twenty has mostly gone, so the remaining constraint is willingness - whether an organisation will actually use the tool it has, rather than defaulting to the multi-month study it already knows how to commission. That’s true as far as it goes. It’s also not the interesting part, and it undersells what’s actually happened to the original advantage.

Here’s the part I missed. What an agentic system does when it “assembles competitive intelligence in minutes” is compress what is already knowable - public filings, trial registries, comparable launches, analyst notes - into a coherent paragraph. It’s extremely good at this, and getting better every quarter. It is not, in any meaningful sense, discovering anything. Compression and insight are not the same thing, even when the output of the first one is fast, fluent and correct.

That distinction didn’t matter much when fast, good compression was itself the scarce resource - when producing a usable answer in an afternoon required an unusual team with unusual pattern recognition and a willingness to say something before it was fully defensible. It matters enormously now that the compression is part of the environment (and in your competitor companies). Point the same tool at the same public data and most teams converge on the same well-shaped answer. What was built to generate asymmetric learning turns into a consensus engine the moment everyone has it. Speed stops being an edge the moment everyone has it - pharma has now watched this happen to more than one capability that looked defensible for about a year.

So the scarce thing was never really the fast answer. It’s the judgement about what a fast answer is actually worth, question by question. Everyone can generate one now. Fewer people can tell whether it’s cleared the bar the decision in front of them requires, or whether it’s just a plausible-sounding restatement of the obvious with the corners smoothed off.

The old failure mode is still around, and AI made it worse before it made it better: the multi-month study, AI-augmented at every stage, that still tests only the one hypothesis the model surfaced first, with better graphics than 2021 could have produced. Treating the model as an oracle rather than a dialogue partner just moves the premature commitment earlier in the process; it doesn’t remove it.

But there’s a newer failure mode running in the other direction, and it’s the one I’d write about if I were writing that post now. It’s stopping at the first coherent answer because it arrived fast and sounded finished. Every one of these moments hides a version of a question computer science has a name for - when do you stop generating options and commit to one? - and the honest answer depends entirely on what the answer needs to be good enough for. Good enough to rank hypotheses is a different bar from good enough to kill a programme, which is a different bar again from good enough to put in front of a regulator.

That’s roughly stage-dependent, in a way the original piece only gestured at. Late-stage work, manufacturing, anything that touches a regulatory filing still wants depth, and AI hasn’t changed the biology or the paperwork. Early commercial strategy and development sequencing - the two- or three-year-out questions about which population, which comparator, which price - remain the zone where the rapid mode is not just adequate but genuinely better than the slow one, exactly as I argued in 2021.

There’s something I didn’t see coming underneath all of this. As compression gets cheap, the things that resist compression become more valuable, not less. Original experimental design. Direct observation of a patient population nobody else has looked at closely. Causal judgement about a mechanism, the kind that only comes from staying with a difficult result past the point where a model would have offered you a tidy explanation and moved on. None of that arrives any faster just because the literature review around it does.

Which is also, I think, why the organisations getting real advantage right now aren’t the ones with the best access to agentic tools - that access is close to universal, and will be fully so within a year or two. They’re the ones with the highest learning velocity: how fast the group’s actual understanding of the molecule improves, as distinct from how fast it can produce a plausible paragraph about it. Smaller teams often do better here, not because they’re smarter, but because they carry less institutional pressure to convert an honest “we don’t know yet” into a defensible three-month process before anyone’s allowed to act on it.

I framed the original piece as a choice between two speeds. I’d frame it now as a set of different bars for how good the answer needs to be, and the actual skill - the one that’s got scarcer while everything around it got faster - is knowing which bar a given question is sitting on before you’ve already committed to an answer.

Speed got cheap. Judgement didn’t.

Five years ago I wrote a short piece pointing out that the industry’s usual response - “multiple shots on goal” - was aimed at the wrong risk. We multiply molecules and hope one survives the same decision process we used last year. The real risk is that the decision process itself is the bottleneck: the way we choose what to learn, how fast we learn it, and whether we ever test alternative ways of deciding.

Looking back at that May 2021 post, two things stand out. First, it landed when this blog was still finding its feet, so almost no one saw it. Second, the core claim has held up as well as I hoped, and better than I honestly expected.

How many companies have genuinely experimented with anything other than a version of the decision process they used the year before? How many run parallel decision systems, or deliberately invite alternative views on an asset’s development path instead of simply hoping they out-thought the competition? 2020 forced temporary flexibility around Covid programmes. Lilly’s approach of keeping acquired biotech groups - Chorus among them - relatively independent still looks like one of the clearer attempts at process diversity - different people and different decision logics running side by side.

The prediction paradigm remains dominant even though we know we cannot reliably predict biology, clinical outcomes or market performance, especially in early phase. We still build elaborate forecasts, compromise product profiles, and then act as if the uncertainty has been removed.

Which brings us back to Flagship Pioneering. In 2021 I noted that if they wanted to launch 100 drugs in ten years they would have to outperform the industry’s best by a factor of three or four, year after year. You cannot do that by hiring smarter people into the same process. You have to behave differently.

Five years later the scorecard is interesting. Flagship has launched something like 142 companies since inception, of which roughly 44 are still active - a ratio that tells you most of what you need to know about how the model actually works. Moderna, the most visible success, has moved from one pandemic vaccine to a multi-product franchise; Acceleron sold to Merck for $11.5 billion (we enjoyed that ride, as I positioned Reblozyl); Sana topped Moderna’s own IPO record. Alongside that: Kaleido shut down in 2022 after burning through a $75 million IPO, Ohana closed eighteen months after launch, Rubius folded in 2023 after ‘losing’ three-quarters of its staff trying to pivot. None of that is scandalous - it’s what a portfolio of speculative platform bets looks like - but it is a long way from a steady stream of novel approved drugs at the rate the original ambition (and statement) implied. The more interesting recent move is the 2023 Pfizer tie-up, $50 million each into ten single-asset programmes structured explicitly to test a new decision model rather than just fund more molecules - a more direct answer to the original question than the 100-drugs number ever was.

That gap between ambition and delivery is useful. Even the most deliberate attempt to operate differently still collides with the same biological, regulatory and translational realities that constrain everyone else. The lesson is not that the ambition was wrong; it is that changing the decision architecture is harder - and more necessary - than most of us admit.

Lilly’s scorecard is harder to read cleanly, if only because it’s impossible to ignore: the company is now the most valuable pharmaceutical business in the world, more than a trillion dollars in (today’s) market value, built substantially on a GLP-1 franchise that Chorus-style units had a hand in de-risking early (fast, agile). Correlation isn’t proof - the biology’s own turn toward obesity as a mainstream indication, and no small amount of timing, did real work, a lot of more obvious decisioning, too - but the five years gave that structural bet every chance to be a liability, and it wasn’t one.

The intervening years have given us new tools that should make asymmetric learning easier. It’s been quite the five years. AI can explore chemical and biological space at scales that were previously impossible. Adaptive trials and real-world data loops can shorten the time between hypothesis and insight. Yet many organisations have simply bolted these tools onto the old prediction machinery. The asymmetry only appears when the organisation is prepared to treat learning itself as the competitive variable - maximising insight per unit of resource, running multiple decision logics, and accepting that most paths will fail while a few teach disproportionately.

So the question remains the same, only sharper. Are we still mostly multiplying molecules inside a single decision process, or are we finally willing to put multiple shots on the goal of how we decide? The companies that treat process diversity and learning velocity as design parameters, rather than afterthoughts, will keep creating the only real edge available in this industry.

The original post was written at the start of this blog. The argument feels more relevant now than it did then. The data has moved a little. The decision problem has not.

People usually aim their line at the top of the org chart - the CEO promising the board that next quarter will turn, the salesperson calling a deal “very likely” before anything is signed.

In pharma, the same mistake often happens much earlier, much lower down, and under the cover of technical seriousness. It happens in trial design. And because it happens in trial design, it wears the costume of scientific rigour.

Call it symmetric learning: running a development programme that teaches you nothing your competitors would not also learn.

That is a harsher definition than most teams would use. But it gets at something real. A lot of programmes are called de-risked when what they really are is standardised, familiar, and unlikely to produce any differentiated understanding.

What makes learning symmetric

Asymmetric learning means structuring work so that you learn faster, more honestly, and more cheaply than the organisation next door, from the same or even less activity.

Symmetric learning is its mirror image, and it is much more common.

It is what happens when your endpoints are the class’s endpoints, your comparator is the class’s comparator, your population is the class’s population, and your Phase 3 is designed - deliberately, and often proudly - to look like the accepted template for the indication. Ask the team why, and the answer is usually some version of the same word: de-risked.

That word sounds disciplined. Often it just means nobody wants to have the argument.

A trial designed to look exactly like the ones that came before it cannot, structurally, produce much information your competitors do not have the moment the data read out. You have not necessarily reduced risk. You may simply have reduced the chance of learning anything distinctive. Those are not the same thing, even if the industry often talks as if they are.

The hope hiding inside “de-risked”

This is the uncomfortable part.

If a programme is designed to generate no differentiated information, then the thing that has to save it is the molecule itself simply being better than the base rate would suggest. Not better because you found where it works, in whom, or under what conditions. Just better by luck, or at least by hope.

At that point, you are not really designing an experiment in the strongest sense. You are paying a great deal of money to run a familiar protocol and hoping reality is unusually kind to you.

That, to me, is the core of symmetric learning: hope in the shape of a development plan.

And it is a peculiar kind of hope. It is not even bold hope. It is not the sort of optimism that backs a strange hypothesis or follows an unexpected early signal somewhere interesting. It is the quieter hope that if you do what everybody else does, your asset will somehow come out better than average anyway.

You are hoping for an edge while designing away the conditions under which an edge could be discovered.

Where it comes from

Almost nobody sets out to build a symmetric programme.

It accumulates the same way a lot of mediocre strategy does: through a series of individually reasonable decisions that add up to an unreasonable whole.

Biostats picks the established endpoint because it is precedented and easier to defend. Regulatory picks the familiar comparator because it is cleaner with agencies. Commercial picks the standard population because the market model already exists. Clinical operations would rather not introduce a design wrinkle that makes execution harder. None of these decisions is absurd on its own. That is exactly the problem.

Each step defers to what has already been learned - usually by somebody else, on somebody else’s molecule, under somebody else’s constraints. The end result is a programme that is easy to explain in a governance meeting and very unlikely to surprise anyone, including you, in a useful way.

The tell

There is a simple diagnostic question here, and it works because people tend to feel the problem as soon as they hear it:

If this trial comes back exactly as clean as we hope, will we have learned anything a competitor running the class-standard design would not already know?

If the honest answer is no, you may not be running much of a learning programme at all. You may just be running a very expensive waiting period with better monitoring.

That is why this matters. The issue is not only whether a trial is well run. It is whether it is capable of generating knowledge that changes your position.

This is a portfolio problem, not just a trial problem

A symmetric trial can absolutely succeed.

That is one reason the pattern survives. A positive readout here, a tolerable outcome there, and the whole model keeps reproducing itself. The reinforcement is intermittent, which makes it stubborn. But the real cost shows up over time, at portfolio level.

An organisation that repeatedly runs symmetric programmes is not necessarily making foolish decisions on any single asset. It is making a broader strategic choice: not to be the one that learns something first. Not to be the one that finds the sharper subgroup, the more useful biomarker, the better sequence, the more revealing failure. And then, a few years later, it wonders why even its positive assets look only modestly differentiated, and Commercial struggle to differentiate.

That should not be surprising. If you repeatedly design for familiarity, you should expect familiar outcomes.

What the alternative actually asks of people

Asymmetric learning is not free, and it is definitely not more comfortable.

It asks teams to hold a hypothesis specific enough to be wrong. It asks them to run the early, cheaper experiment that could embarrass them, instead of the safer one that merely keeps everyone calm. It asks them to walk into governance with a design that prompts the question, “Why aren’t we doing it the usual way?” and to have a better answer than “because this feels more innovative.”

That is a cultural demand as much as a scientific one.

Symmetric learning asks the organisation to feel safe. Asymmetric learning asks it to find something out. Only one of those is a strategy.

A practical test

Before the next protocol is finalised, ask the team to name one thing this trial could show that a competitor running the class-standard design would not learn.

Not in ten slides. Not after a week of backfilling. In the room.

If nobody can answer clearly and quickly, there is a good chance you have not designed an experiment so much as a hope - and scheduled it eighteen months into the future.

These are not temporary inconveniences. They are structural. This is classic wicked-problem territory - the kind Rittel and Webber described more than fifty years ago. Wicked problems have no definitive formulation (the way you frame the issue already shapes the possible answers), no clean stopping rule that tells you when you are done, and no true-or-false solutions, only ones that are better or worse depending on whose values you prioritise. Every intervention is essentially a one-shot operation with consequences that are hard or impossible to reverse, and each case is unique enough that yesterday’s playbook rarely transfers cleanly. Incomplete and contradictory information is the normal state, not an exception.

The necessity is not in doubt. What still feels scarce is the kind of invention that actually matches it.

Most organisations still respond with some version of the old symmetric playbook: fixed target product profiles (or compromise product profiles) written too early, stage-gates that assume the map is already there, development plans built around precedent, and review meetings that treat deviation from the original plan as the greater risk. It all sounds disciplined. Often, it is just familiar.

That is parrot culture.

Parrots are excellent at replication. They learn the language of the last successful programme, repeat the accepted forms of rigour, and get very good at looking competent inside the existing system. In a tame world, that can be efficient. In a wicked one, it becomes a liability. Everyone studies the same incomplete charts and reaches the same expensive dead ends - only a bit later.

Pirates operate differently, me hearties.

They do not wait for a perfect map because they know they are not getting one. They move toward the high-uncertainty, high-upside water, take cheap probes, gather asymmetric information, and redraw the chart as they go. They treat early signals, failed hypotheses, unexpected biomarkers and staged collaborations as learning currency. They assume the first version of the plan will be wrong, and they design the organisation so that being wrong is cheap, fast and informative.

That is asymmetric learning in practice.

Give two teams the same molecule. The parrot team predicts, executes, and hopes the later studies confirm the prediction. The pirate team designs for optionality - biomarker splits, adaptive elements, early signal mining, staged external bets that turn outside knowledge into internal edge.

If the data cooperate, the advantage compounds. If they do not, the pirate team kills or pivots earlier, with better information and less attachment to the original story. Over a single asset that may just look like style. Across a portfolio, it becomes structure.

You can already see versions of this across the industry.

• Gilead’s multi-year staged collaboration with Arcellx on anito-cel generated the clinical and operational insights that others lacked. When the data aligned, Gilead moved to full acquisition in early 2026.

• Lantern Pharma used AI to mine noisy Phase 1a signals from LP-184, identified PTGR1 and DDR biomarkers, and converted a single dose-escalation study into multiple parallel, biomarker-guided expansion paths rather than a conventional broad Phase 2.

• Insilico’s AI-generated Rentosertib progressed from novel target nomination through positive Phase IIa data into Phase III for idiopathic pulmonary fibrosis within a compressed timeline, while the broader pipeline continued to spawn new options.

Those are pirate moves. Not anti-rigour. Not chaos. Just rigour applied to the fact that the problem is uncertain in the first place.

Culture decides which behaviour survives.

If your review meetings punish deviation from the original plan more harshly than they reward new insight, you are going to get parrots.

If early failure is treated mainly as career risk, people will avoid pirate behaviour no matter what the strategy deck says.

If “we’ve always done it this way” still counts as a serious argument, the map stays fixed while the territory moves.

So the practical question is simple: what would your organisation have to change to make pirate behaviour the safer, more rewarded path?

That might mean parallel teams on the same asset. It might mean explicit permission - and budget - for high-optionality early experiments. It might mean changing success metrics so that the quality and speed of learning count alongside the binary go/no-go. It might mean treating the prediction paradigm itself as the hypothesis under test, rather than the unexamined default.

We’ve got the necessity. One of the few durable edges left is building systems that can learn from it faster, more honestly, and more asymmetrically than the competition.

Parrots will keep reciting the old charts.
Pirates will keep redrawing them.

It’s available for viewing/ download here.

I obviously don’t know if it’s ready for prime time yet (or for LinkedIn 😅), so I’d welcome any and all comments on what’s here…

It is a fight that matters a lot. To put a number on why either side is bothering: Mounjaro and Zepbound - the drugs that retatrutide is designed to succeed - brought Lilly $36.5 billion in 2025 alone, in a single year, from a single company. Retatrutide doesn't need to match that for this fight to be worth having. It needs a fraction of it, protected for a few extra years, to be worth billions. And exclusivity isn't even the only clock tied to this word: peptide. Under the Inflation Reduction Act, small-molecule drugs become eligible for Medicare price negotiation 9 years after approval; biologics get 13. Winning biologic status doesn't just buy Lilly 7 additional years of exclusivity - it (currently) buys 4 additional years before the federal government can start negotiating the price down at all. Two separate countdowns, both of them affected by the same classification.

The fight is over a definition, not a drug. That matters because definitions are where asymmetric advantage often hides.

Retatrutide has 41 amino acids. But the real dispute is not the number itself - everyone can see that. It is which amino acids count, under what definition, and who gets to decide. On that question, Eli Lilly and the FDA have been in open legal conflict since September 2024. A federal court has already given its view. The case is now on appeal. And Lilly’s target submission window in early 2027 is approaching with the classification still unresolved.

The regulatory mechanics are simple enough. FDA regulation draws a line at 40 amino acids. Cross it, and a molecule can qualify as a “protein”, making it eligible to be regulated as a biological product - a Biologics License Application, with 12 years of exclusivity. Stay under it, and you are filing a New Drug Application instead, with five. Same molecule, broadly similar clinical package, largely similar manufacturing realities, same patients. Different piece of paper, different decade of market protection.

Lilly’s argument is that retatrutide should be counted using total amino acids, which puts it at 41 and over the line. The FDA’s argument has been that only “alpha amino acids” count - a narrower category that keeps retatrutide below the threshold and off the biologic pathway.

In September 2025, the Southern District of Indiana ruled on part of that dispute. The court sided with the FDA on the narrow question: retatrutide does not satisfy the strict statutory definition of a “protein” under the 40-alpha-amino-acid test. But the court rejected the agency’s fallback argument that retatrutide is also not “analogous to a protein”, even though the statute allows that second category. On that point, the court said that the FDA’s reasoning “flouts the statutory text and sidesteps congressional intent”. So the agency won the narrower definitional fight but lost the broader interpretive one.

That is the interesting part. The FDA was sent back to do something it had never properly done: define what “analogous to a protein” actually means, instead of smuggling the same numerical threshold back in under a different label. Lilly has since appealed, seeking a fuller, more declarative win. The FDA still has to articulate a definition it does not yet seem to have, for a category it has been using for years without clearly spelling it out. Timing may force a resolution before filing. It may not.

Nobody in this dispute is arguing about what retatrutide does in a patient. The pharmacology is not really in question. What is in question is which regulatory bucket the molecule belongs in. The bucket was formalised in 2020, building on earlier guidance, to create administrative clarity as peptide and protein complexity increased. Retatrutide did not break the rule. It just sat close enough to the boundary to force a more uncomfortable question: what was the rule actually for?

That shape should feel familiar. I wrote a few months ago, in “The Looking-Glass Moment in Pharma”, that the industry’s real advantage is shifting from predicting better to learning faster - from committing early to fixed assumptions towards structured adaptation as new information arrives. Regulatory categories are usually treated as the one stable part of the system. The science moves. The market moves. The rulebook does not. Retatrutide is a reminder that the rulebook moves too. It just moves more slowly, and usually only when someone pushes hard enough to make the people defending it explain themselves in court.

This is not even the first time the line between “drug” and “biologic” has shifted in commercially meaningful ways. In 2020, a whole class of long-approved protein products - insulin, human growth hormone, and several reproductive hormones - was formally reclassified from NDAs to BLAs under the BPCI Act’s “deemed to be a licence” provision. That transition cut in the opposite commercial direction from the one Lilly wants here. It was not companies fighting for a longer exclusivity runway. It was regulation closing a loophole so that older insulin products would finally face biosimilar competition.

The underlying move, though, was the same: a molecule crossing from one regulatory category into another. The commercial consequences depended entirely on where in the product life cycle that crossing happened, and to whom. In that narrow sense, the category is commercially neutral. Whether you want to be called a biologic depends on where you are standing when the definition changes. Of course, the surrounding regime is not truly neutral - inspection standards, interchangeability rules, and post-approval pathways all matter - but the central point is still there.

That is the bigger lesson. Every “fixed” boundary in this industry - drug vs biologic, on-label vs off-label, generic-eligible vs not - was drawn by someone, at some point, using the science and administrative priorities available then. Yet those boundaries go on being treated as fixed long after the underlying molecules have outgrown them. Most companies wait for a court or Congress to force a redefinition, as happened with insulin. A smaller number go looking for the edge deliberately, as Lilly appears to be doing now, because they have noticed that a definition nobody has stress-tested against engineered multi-agonists in twenty years is not a fixed wall. It is a hypothesis.

You can already see what that implies for the next design cycle: peptides engineered to sit just over or just under the line depending on the desired pathway; sponsors testing how far “analogous” can stretch once the FDA is forced to write the criteria down; regulatory strategy becoming part of molecular design rather than something bolted on afterwards. Europe’s more process-oriented approach to similar molecules is a useful reminder that the US line count is a policy choice, not a scientific necessity.

Question the molecule less. Question the definition more.

This isn’t a philosophical indulgence (I am allowed those, but this is not one). It’s the single most consequential question a portfolio function can ask about itself, and most never ask it properly - because the answer is uncomfortable, and because the industry’s feedback loops are too slow and too thin to answer it honestly.

The problem is sample size, not judgement

Most pickers in pharma - portfolio leads, BD heads, even whole R&D organisations - place a small number of large, irreversible bets across a career. That’s the wrong regime for learning who’s good. A poker player sees thousands of hands and can, eventually, separate skill from variance through sheer volume. A pharma executive might make a handful of genuinely consequential calls in twenty years. At that frequency, a couple of lucky breaks look indistinguishable from a repeatable edge - and a couple of bad breaks can end a career that was, in process terms, sound all along.

This is worth saying plainly because the industry’s storytelling runs the other way. We build hero narratives around the person who picked the asset (or worse, just ‘invented’ the asset) that became the blockbuster, and we rarely ask what the ex-ante odds were, or how many similar bets they made that quietly failed. Survivorship dresses itself up as talent.

One good pick is a data point, not a track record. The more useful test is whether someone is right more often than base rates predict, across genuinely different situations - different modalities, different disease areas, different competitive landscapes. A single blockbuster tells you almost nothing about the picker. A portfolio of thirty assets, de-risked and licensed across a spread of therapeutic areas with a return rate that beats the field, starts to tell you something real. This is the difference between celebrating one bet that landed and trusting a process that keeps landing bets - closer to Novartis’ current portfolio approach than to any single-asset story.

Resulting: the trap that makes it worse

Poker players have a word for the mistake of judging a decision by its outcome rather than its quality at the time it was made: resulting. (You can check my interview with Annie Duke on YouTube.) It’s endemic in pharma. A drug that limps through Phase III on a mechanism nobody fully understood gets remembered as a brilliant pick. A drug that fails despite a well-built, evidence-updating development plan gets remembered as a bad one. Both judgements can be wrong.

The alternative is to separate two questions that we habitually collapse into one:

• Was the outcome good?

• Was the decision good, given what was knowable at the time?

A good decision can produce a bad outcome, and a bad decision can produce a good one. Confusing the two is how organisations end up promoting the lucky and firing the unlucky, while the actual skill - the thing worth retaining and replicating - goes unmeasured in either direction.

Bad bets that do work out might still result in a product that’s hard to market, Many subpar launches show that bad products do make it to market, and they underwhelm. I have had to reposition drugs that launched badly (they’ve turned around well - positioning really does have a tangible, visible outcome when done well). But that does point to the underlying question - at what point would you call a decision good, or lucky?

A test: does the pick have a story that could have failed?

If resulting is the trap, here’s a rough diagnostic for climbing out of it. Ask, of any pick, whether the picker could have told you in advance:

• what specific evidence would prove them wrong

• what threshold would trigger a kill or a pivot

• what they expected to learn from the next piece of data, before they saw it

A good example might sound like this: “If the Phase 2b primary misses by more than X and the key biomarker subgroup shows no signal, we kill. If safety is manageable and the secondary functional score moves, we expand into indication Y and re-size the Phase 3.” That story existed before the result. Luck has no such story. It is a pick defended after the fact with whatever evidence happens to be lying around.

Good picking has a story that existed before the result did, and one that would have changed the picker’s mind under a different, equally plausible set of facts. This is really the same idea I’ve been circling in recent posts under the banner of asymmetric learning: the value isn’t in being right, it’s in building a process that updates faster and more honestly than the next team’s.

It’s also why the current wave of AI-native discovery platforms is interesting for reasons beyond speed. Faster, cheaper trials don’t just compress timelines - they compress the feedback loop that lets you tell luck from skill apart in the first place. If you can get a real signal in months instead of years, pre-registering what would change your mind becomes cheaper and more frequent. You get more shots at learning which of your people (or your models) are actually calibrated, rather than just occasionally, expensively, right.

Why this matters more now, not less

As development gets faster and cheaper to test, the temptation will be to declare victory earlier - to mistake an early positive signal for proof of picking skill, rather than as one more data point in a track record still being built. The organisations that get this right won’t be the ones with the best individual calls. They’ll be the ones that have built the infrastructure - fast trials, honest pre-registration of what would change their mind, real memory of near-misses as well as wins - to tell the difference between a good decision and a good result, and to keep score on the right one.

Of course, biotech is a harder thing to spot. Their decisions tend to fall into guided luck by implication - they don’t make enough to spot a pattern. But that’s not a great story for investors - they’ll carefully present the evidence as if there’s a simple logic, and will be tempted to leave out the things they don’t know.

Picking well is a discipline you can build. Picking luckily is a story you tell afterwards. The whole point of asymmetric learning is to make sure your organisation can’t confuse the two.

Pharma has spent decades pretending otherwise.

Look at almost any early development programme and the same double-think appears almost immediately.

• We will define a single, detailed Target Product Profile early and retain real strategic optionality later.

• We will run increasingly large, decision-shaping studies while insisting we are still merely “learning”.

• We will demand predictive confidence from Phase I/II and still call the process exploratory.

• We will keep every programme alive and somehow concentrate resources only on the genuine winners.

These are not complementary ambitions. They are trade-offs. Trying to hold both sides at once is simply the corporate form of wanting to eat the cake and still present it intact at the next governance meeting.

The rigid TPP is the cleanest expression of the problem. It is certainty borrowed against a future nobody can see. It assumes that, years in advance, we can specify the profile that will prove both approvable and commercially decisive in a market that does not yet exist. That is rarely strategy. More often it is compromise product profile (CPP) dressed up in development language: a prediction presented as a plan.

The effect is subtle but damaging. Early development stops being a search for high-value information and becomes an exercise in defending a preferred future. Data that fit the original frame are amplified. Data that threaten it are explained away, deferred, or quietly absorbed into a revised story. The TPP is not really abandoned. It is rewritten - and then treated as though it had always been more flexible than it was.

That is not learning. It is narrative maintenance with a governance deck.

What asymmetric learning actually demands

Asymmetric learning begins from a less comfortable premise: in this industry, outcomes are lopsided. Most assets fail. The few that succeed often do so for reasons that were only partly visible at the outset, or not visible at all. If that is true, then the rational early strategy is not to defend one preferred future. It is to maximise the chance of discovering a better one before everyone else does.

That requires explicit choices.

You cannot preserve every option and still move quickly. You cannot demand certainty from noisy early data and still claim to be exploring. You cannot keep protecting sunk-cost stories and also free capital for the next set of high-upside probes.

There is no magical third way here. The firms that create outsized value are usually not the ones that found a way to reconcile the irreconcilable. They are the ones that chose early, before the data made the choice for them - or before their internal politics made the choice impossible to admit.

What that looks like in practice

Semaglutide is a useful example. Novo Nordisk had a diabetes asset. It also had an obesity signal that could easily have been treated as secondary, messy, or strategically inconvenient. Instead, that signal was pursued as the more valuable opportunity. Ozempic was approved for type 2 diabetes in 2017; Wegovy, the higher-dose obesity brand, followed in 2021. The point is not just speed. It is willingness to let the more valuable future displace the tidier original one. The category now dwarfs the frame in which the molecule first looked most legible because the off-path signal was treated as the asset, not the distraction.

Pembrolizumab reflects the same logic in another form. Merck did not protect one narrow future for the molecule and then defend it to exhaustion. It built an architecture that allowed value to accumulate across many paths. Basket trials were not merely a clinical design choice. They were a strategic refusal to collapse optionality too early.

Sildenafil is cleaner still. As an angina compound, it disappointed. As a response to an unexpected but commercially decisive signal, it became Viagra. That was not the reward for clinging harder to the original plan. It was the reward for recognising that the plan itself had become the wrong asset.

In each case, value came not from confirming the first story, but from being willing to replace it.

The cost of refusing the trade-off

The industry’s default response is to avoid the choice for as long as possible. The result is familiar. Pipelines fill with programmes that are neither clearly alive nor clearly dead. Teams generate data optimised to support the inherited story rather than to answer the questions that would genuinely change the asset’s trajectory. Late-stage failures become more expensive because nobody was willing to take the earlier loss. Portfolio reviews turn into exercises in rhetorical salvage.

This is what symmetric learning looks like in practice: similar companies, running similar programmes, generating similar information at similar cost, on roughly the same timetable. Everyone claims to value optionality. In reality, most organisations treat it as a threat the moment it destabilises the preferred narrative.

That is the deeper confusion. Optionality is not the thing early development should sacrifice in order to preserve discipline. In many cases, it is the thing discipline should be preserving. The rigid TPP does not solve that tension. It simply allows the organisation to postpone admitting that a choice is being made at all.

A practical test

The next time you are in an asset review or portfolio discussion, ask the room a simple question: are we trying to protect the current preferred future, or maximise the chance of finding a better one?

If the honest answer is “both”, the discussion is already off course.

At that point, you are not managing a learning system. You are managing a rewritten TPP - one that no longer describes where the asset is going, only what the organisation is still unwilling to admit about where its original prediction failed.

I’ll credit Substack - it surfaced this original post for me, and suggested I might want to update it - it was a useful prompt…

Pharma still runs on a model I’ll call Predict, Pick, Plan. You predict the outcome of a programme - usually by attaching an eNPV to a single indication and a single development path. You pick the path with the best number. You plan a linear route to get there: target, mechanism, indication, trial design, launch, all committed to in roughly that order, roughly all at once.

The problem was never that the predictions are bad. The problem is that the entire exercise is structurally incapable of being good, no matter how sharp the people running it are.

Why “predict” can’t be fixed

An eNPV is only ever a statement about one path. Run the numbers on the indication you’ve chosen, and you’ve said nothing about the indications you didn’t choose - you’ve simply declined to look. Average across several plausible paths instead, and you’ve built a number that describes no version of reality anyone will actually live through. There is no third option where the eNPV is both meaningful and comprehensive. You can be precise about one future or vague about many; you cannot be precise about many.

That would be a survivable flaw if the early estimates feeding the calculation were even roughly right. They aren’t, and they can’t be - not because pharma forecasters are bad at their jobs, but because the confidence intervals on preclinical and early-clinical assumptions are wide enough to swallow the conclusion. You are not predicting the outcome. You are dressing up a guess in a discounted cash flow model and then treating the output with a confidence the inputs never earned.

Four years on, I’d add a sharper version of this point, borrowed from a piece I wrote this year on what pharma could learn from how machine learning teams actually spend their time: the ceiling on a program isn’t set by the trial that reads out, it’s set by the categories and assumptions you fixed before you’d learned anything. Bad early framing - the wrong endpoint, the wrong subgroup, the wrong idea of what the drug is for - creates a noise floor no amount of statistical power later can lift you above. PPP doesn’t just make a bad prediction. It makes a bad prediction and then spends years refusing to revisit the frame that produced it.

Why “pick” doesn’t work either

Picking assumes someone, somewhere, is good enough at this to be worth listening to. They aren’t, and the industry now has better data than it did in 2022 confirming it. Phase I-to-approval success sits at roughly 7% as of 2024 - barely moved from 6% the year before, and still below the 9% the industry was managing in 2015-2019. Our diagnostics have improved. Our information is better than it’s ever been. Pick a programme today and you are, on average, worse positioned to be right than a team was a decade ago. That’s not a data problem. It’s a process problem wearing a data costume.

Why "plan" locks in the damage

Here’s the part I still think gets underrated: even a correct pick loses most of its value once it’s forced through a rigid plan. The plan gets made at Predict - the single moment you know the least you’ll ever know. Everything that happens next is evidence: a strong read, a subgroup signal, a side effect nobody expected. But the plan wasn’t built to use that evidence, because using it would mean admitting the original path might be wrong. So committing early isn’t a risk you take once. It’s a decision you keep making - every quarter - not to look at what you’re learning.

What’s changed since 2022 - the world caught up a bit

I didn’t expect to be writing this next section four years ago, but here we are: some of the biggest institutions in the industry have started acting on a version of this argument, whether they’d put it in these terms or not.

In February 2026, FDA Commissioner Marty Makary and CBER’s Vinay Prasad announced in the New England Journal of Medicine that a single well-designed pivotal trial, backed by confirmatory evidence, is now the default standard for approval - replacing a two-trial requirement that had stood since the 1960s. Their stated reasoning was that modern trial design and statistical methods establish credibility in more ways than duplicating the same study twice. Read plainly, that’s a regulator telling the industry that the quality and structure of evidence matters more than ritual repetition of a fixed plan - which is close to the whole argument of this post, said from the other side of the review desk. I wrote about the wider shift this signals in “The Looking-Glass Moment in Pharma” - worth a read if you want the fuller regulatory picture, because the FDA pilot behind this went further than the headline: shorter feedback loops, lower cost of early failure, and an explicit incentive to test more than one path before committing hard to any of them.

There’s also a return-on-investment story that didn’t exist when I wrote the original piece. Deloitte’s latest pharma innovation report shows industry R&D returns rising for a second straight year, to around 7% - a real if partial reversal of the decline that was the backdrop to my 2022 argument. Most of that improvement traces to GLP-1s. That’s not a coincidence worth driving past on the motorway: the value in GLP-1s didn’t come from a well-predicted single path chosen early. It came from breadth - a diabetes program that turned out to have an obesity opportunity hiding inside it, spotted late, by people willing to keep looking after the original plan said they were done. I’ve written elsewhere about just how much earlier that value could have been unlocked if “what if” thinking had been built into the process rather than left to a handful of persistent believers - and, separately, about my own record of being wrong on the timing of an industry shift while still being right about its direction, which is exactly the asymmetry PPP-style planning can’t accommodate. A rigid plan doesn’t just get things wrong; it gets them wrong in a way that forecloses being usefully wrong.

The same trap, wearing new clothes

Here’s the part that worries me more than anything above, and it’s genuinely new: AI is now industrialising exactly the two stages of PPP that were always the least of the problem. As of early 2026 there are roughly 173 AI-discovered programs in clinical development, with early-phase success rates jumping to somewhere in the 80-90% range against a historical baseline near 52%. That sounds like the prediction problem is solved. It isn’t. Only 15-20 of those programmes were expected to reach Phase III this year - because AI has made candidates cheaper and faster to generate, without making them any cheaper or faster to kill, and without touching the part of drug development that was always the actual bottleneck: proving efficacy in real patients. Pump more candidates into a Predict-Pick-Plan pipeline at higher velocity and you don’t get a better pipeline. You get the same trap, better funded, moving faster toward the same wall - unless the extra throughput is paired with real optionality and real kill discipline further downstream, which is exactly what PPP-style planning was never built to provide.

What to do instead

The alternative I proposed in 2022 was an asymmetric learning approach: hold multiple destinations open, design early experiments for what they teach you rather than what they prove, and specialize late - treating the choice of a single path as something to be earned by evidence, not assumed at the outset. I’d still stand behind every part of that. What I’d add now is a more concrete account of how, because I’ve spent the years since trying to answer that question properly (and working with clients trying to) rather than just naming it.

Five things seem to matter in practice. Search - deliberately generating more candidate paths than intuition alone would surface, and treating that breadth as an asset rather than an inefficiency. Feedback - extracting information from every experiment, including the ones that “fail,” rather than treating trials purely as pass/fail gates. Infrastructure - the unglamorous work of data standards and ontology, because a program built on crude categories can’t learn its way out of a bad frame no matter how good the science is. Competition - running live comparisons between hypotheses instead of letting the first plausible narrative harden into consensus by default. And human judgment, redirected rather than removed - toward framing the right questions and catching flawed proxies, which is the one part of this no model does for you.

None of that is a plan in the PPP sense. It’s a decision system: a way of weighing the cost of finding something out against the value of knowing it, repeated at every stage, rather than answered once at the start and defended forever after.

Where this leads

This post turned out to be more of a beginning than I realised when I wrote it. Most of what this newsletter has become since - the case for treating development as a discipline rather than a forecast, the argument that winning is a process and not a gift for prediction - traces back to the question this piece was really asking: if you handed the same molecule to two different companies, would it end up in the same place? In 2022 I didn’t think so. I still don’t. The difference now is I have a better answer for what to do about it.

For the fuller picture: “Picking Winners (Updated 2026)” makes the case that no one can reliably pick winners in the first place; “The Looking-Glass Moment in Pharma” covers the regulatory shift in more depth; and “Asymmetric Learning: What This Blog Has Been Trying to Say (All Along)” is the closest thing I have to a mission statement for everything downstream of this post.

So if anything the direct number is conservative for the part of the market that matters most. But the argument I want to make doesn’t depend on it, which is why I’m happy to give it away at the top. Forty per cent or fifteen, something has changed in the shape of the business, and the change is more interesting than the share.

Most of the commentary has stayed on commercial ground. Cutting out PBMs, cash pricing, competing with compounders, simplifying access. All true, all useful. The quieter shift is structural.

Manufacturers have mostly lived downstream of their own products. Prescribing behaviour, real persistence, dose titration in the wild, who drops off and when - these signals arrived late, filtered and shared. You waited for the claims data, the syndicated reports, the secondary analyses, and by the time you had a picture everyone else had a version of it too.

Mostly, but not entirely. Rare disease and specialty have run hubs and patient support programmes for years, and those functions have sat close to the patient for a long time. It’s worth saying, because it means we aren’t speculating about whether the mechanism works. We know it does. What’s new is the scale: the same proximity applied to a category with tens of millions of patients rather than a few thousand.

When a meaningful share of volume runs through your own channel, the lag collapses. You see the market closer to the moment it happens. Not perfectly and not completely, but earlier, and with a fidelity the traditional distribution system cannot match. What used to be a shared lagging indicator becomes a private leading one.

The commercial advantage is obvious enough. The longer-term effect is subtler. Future assets launch into a different information environment, because you already know something about how patients use the class, where the friction sits, which segments persist. That shapes what you optimise for next - formulation, support services, the design of the clinical programme itself. The company that sees the real-world patterns first gets to decide what “next” looks like while everyone else is still reconciling last quarter’s claims.

The obvious objection, and I don’t think it has a clean answer, is that the people who arrive through a direct channel are not a random sample. In obesity this is subtler than it first looks, because cash-pay isn’t some fringe of the market. It’s close to the centre of it. But the patient who can find several hundred dollars a month is still not the patient who can’t, and direct-channel populations skew to the younger and wealthier, telehealth-acquired, lighter on comorbidity, more likely to be treating something nearer to appearance than disease. The proprietary leading indicator leads on a self-selected group, and the speed with which you learn about that group can look a great deal like knowing the market.

There’s a related loss. Volume that moves out of the traditional system stops generating the signals you used to buy, expensively. You gain resolution on your own patients and give up a little on everyone else’s.

None of that is a reason to dismiss the advantage. It’s a reason to be careful about what kind of advantage it is. Fast, biased data is a difficult thing to hold well, more dangerous in some ways than slow, representative data, because the confidence it produces doesn’t announce its own limits. The companies that get this right will be the ones treating their channel as one instrument among several rather than the definitive read.

There’s a second constraint, less discussed and probably more binding. Collecting the data is the easy part. Whether anyone in R&D is permitted to see it is another question entirely. Commercial operations and development sit under different governance, and the consent basis on which a patient signs up to a direct pharmacy is not obviously the basis on which you feed their behaviour into a development decision. Most organisations will find that the pipe exists and the valve is shut, and that opening it is a legal and structural problem rather than a data one.

I’ve written before about the movement from molecule to product. What’s happening now is the step after that, from product to a learning system built around the product.

Readers here will recognise the shape of it. The argument I keep making about development is that the same asset handed to two different teams produces two different outcomes, and the gap is almost never talent or budget. It’s that one team planned to learn and the other planned to prove. One treated the programme as a way to know something the competition didn’t. The other ran the studies it was told to run and found out what it already suspected.

What’s interesting about the direct channel is that this is the same argument arriving at the far end of the lifecycle, where it has been conspicuously absent. Commercial is usually where organisations stop learning and start executing. The plan is set, the launch happens, the numbers come in, and such learning as occurs is retrospective and shared with everyone who buys the same reports. Owning the channel changes what’s possible there. Whether it changes what actually happens is a separate question.

Because a direct channel is a learning design decision dressed up as a distribution decision. Most companies will take it as a distribution decision - margin, access, competing with compounders, all the things the commentary is currently about - and will get precisely what a distribution decision gets you. The ones who take it as a learning decision will start somewhere else entirely: not with how much volume we can route through this, but with what we want to find out, and what we’d have to build in order to find it out first.

Two companies can put identical assets through identical channels and end up in different places. They always could. The channel simply raises the ceiling on how different.

So I don’t think the interesting question is who builds the pharmacy. Everybody will build the pharmacy. The question is who has an organisation capable of being changed by what it learns there. That’s rarer than a distribution strategy and far harder to buy, because it asks you to be willing to discover you were wrong about something you’d already signed off.

The slosh was a feature, not a bug, in fairness - they were aiming for smoothness. The fluid coupling smoothed the shifts and saved the engine (ever bigger, brute force, glorious V8s…) the abrupt argument a manual clutch has with a gearbox. But you always paid for the comfort in response - foot down, the engine roared (a glorious V8 roar…), and a beat passed before the car remembered what you had asked of it. Detroit sold that as refinement for forty years, up until the Germans and the dual-clutch showed everyone what a transmission felt like when it was actually connected to something, and “slushbox” became the punchline it still is.

I keep coming back to it, partly because I’m a car guy, but also because pharma has spent the last few years building itself a genuinely excellent engine, and bolting it to a gearbox that would look at home under a 1974 Buick.

The engine - target ID, molecule design, the whole discovery stack - has had a proper couple of car decades. Insilico Medicine took an AI-designed TNIK inhibitor, rentosertib, from idea to Phase IIa data good enough for Nature Medicine in a fraction of the usual time (and by this month it’s into Phase III) - with an inhaled version already cleared by the FDA for its own trial, reportedly the first fully AI-designed molecule to go straight for the lungs. Exscientia was running eleven-month design cycles with a tenth of the usual compound count before most people took the approach seriously; it’s since been folded into Recursion, because it seems that the fastest way to build a bigger engine is to go and buy someone else’s (as with the car industry). AlphaFold handed the industry the shape of nearly every protein worth having. mRNA went from curiosity to platform in eighteen months flat because a pandemic left it no choice. GLP-1s turned a diabetes side-line into a bigger commercial force than PD-1/PD-L1 inhibitors - oncology's own blockbuster class - inside a few short years.

None of that’s just marketing. The molecules are better, the search space is bigger, the feedback loops are tighter. The engine is running properly hot (or cold - heat is wasted energy…).

And then there’s the gearbox. (I always get distracted by gearboxes - when I got to the first car I loved, way back, it would be running around 3500rpm at 70mph in top gear. Now, many cars will be doing 1200-1300rpm at the same speed, while sipping a lot less petrol per rev.)

A new drug still takes the thick end of a decade and a half from idea to approval. Cost per approved asset hit $2.23 billion last year, according to Deloitte, up again from $2.12 billion the year before - this despite everyone still insisting that new tools would finally bend that curve downward. Fewer than one in ten compounds that reach Phase 1 ever reach a patient. Phase 2 is still where good drugs go to die, at something like a 60 percent failure rate. My friend Bernard Munos spent years documenting that pharma’s output of new drugs per company has barely moved in sixty years, however much money gets thrown at it; Another friend, Jack Scannell, and his co-authors gave the whole grim pattern a name in 2012 - Eroom’s Law (Moore’s backwards, drugs per billion dollars of spend roughly halving every nine years). Two economists, two different routes into the data, the same number staring back at both of them: the industry keeps getting better at making candidates and no better at turning them into medicines.

The bottleneck moved. It used to sit upstream - find the target, find the molecule, survive lead optimisation. We’ve more or less cracked that bit now. It sits downstream instead, and downstream hasn’t shifted much at all. The FDA still wants the same tox packages and the same GMP scale-up timelines whatever software drew the molecule. A clinical ops team can’t recruit a Phase 3 any faster for knowing the crystal structure. The engine roars, but the car surges forward about eighteen months later than you asked it to.

Slosh

There’s a nastier version of this, and it’s the bit that should worry anyone paid to extract commercial value from a pipeline: the slushbox doesn’t just cost you time, it costs you signal. In a real torque converter, the slip between impeller and turbine turns into heat, which is just wasted energy. Downstream in pharma, the wasted energy is information. A cellular phenotype isn’t a disease. A mouse isn’t a patient. Recursion’s REC-994 cleared its safety bar in Phase 2 for cerebral cavernous malformation and then handed back an efficacy signal thin enough to disappoint everyone - not because the molecule was rubbish, but because the model had been optimised for the cell, and the cell isn’t the system. The coupling ate the signal on the way through.

That’s the real asymmetry. Upstream is now a precision instrument. Downstream is still a torque converter built for an era when the engine was the unreliable part, plus everybody wanted the slack.

Hope

Here’s the bit I didn’t expect to be writing eighteen months ago: someone appears to be trying to fit a lock-up clutch.

The FDA’s new Commissioner’s National Priority Voucher scheme promises reviews in one to two months instead of ten, using what Marty Makary calls a “tumour board” approach - a multidisciplinary team working a submission all at once instead of relay-racing it through departments. Nine vouchers went out last October; the first full approval under the scheme, a GSK antibiotic, landed within months. The agency’s also rolled out Elsa, an agency-wide generative AI tool for its own reviewers, which is the loveliest irony going: the regulator trying to fix its own slushbox with the same species of technology that built the better engine upstream, and picking up exactly the same “can we actually trust what this thing just told us” headache for its trouble.

Europe’s having a go too, in its own European way, with the new Joint Clinical Assessment replacing twenty-seven overlapping national HTA reviews with one.

On the clinical side proper, digital twins and AI-generated synthetic control arms are starting to do real work - using a patient’s own data to shrink or replace a control group, with the FDA already at the table on the methodology. And when a ten-month-old in Philadelphia needed a gene-editing therapy built for a mutation nobody else on Earth has, the agency found a “plausible mechanism” pathway and got it done in months rather than years - proof the whole apparatus can move fast when the alternative is watching a baby run out of time, which rather undercuts the idea that regulatory inertia is some fixed law of nature rather than a default setting somebody chose.

Don’t get carried away, mind. The same period also produced Most Favored Nation pricing orders and tariff threats, stacked on top of the payer and HTA friction that was already there - a new kind of commercial slip that has nothing to do with trial design and everything to do with Washington. And industry R&D returns climbing to 5.9 percent sounds like the transmission’s fixed itself, until you notice that’s almost entirely GLP-1s doing the work; strip them out and it’s 3.8 percent, barely off last year’s number. Obesity is carrying the whole “things are improving” headline on its own back. (Remember when many companies tried to exit exactly that area a couple of decades ago, much as they’re trying in neuroscience now…)

Acceleration

The lesson from Detroit is that you don’t fix a slushbox by fitting a bigger engine. The actual lock-up clutch arrived in 1949 - and then took the best part of three decades to become something drivers would trust, by which point the slushbox itself still had another thirty-odd years left to run before anyone seriously killed it off. First attempts at a fix are usually clunky, occasionally wrong, and slower to earn trust than the people who built them would like.

Which is roughly where pharma’s new gearbox sits today. Voucher pilots. An AI reviewer nobody’s quite sure they trust yet. Digital twins doing their first proper trial work. A regulatory pathway invented for exactly one baby. It’s not a finished transmission. It’s 1949.

The engine's better than it's ever been. Nobody's short of horsepower any more, and nobody's going to stop building more powerful engines (fingers crossed) - that trade isn't dying, it's just stopped being where the game's won. The game's the clutch now: who gets it to lock up reliably first, and who's still explaining to the board why the car surges forward half a second after everyone's already put their foot down.

Clearly, AI is no longer just a promising tool in drug discovery - it is actively redesigning foundational technologies, enabling new experimental paradigms, and earning serious investment from Big Pharma. Even one week (last week) delivered several standout developments that highlight how AI is accelerating and transforming pharmaceutical R&D. If you’re not excited, you’re not paying attention.

From Nobel laureate-led breakthroughs in gene editing to massive funding for AI antibody design platforms, here are the stories (I think are) worth watching.

1. Doudna Lab Leverages AI to Engineer a Next-Gen Gene Editor

Jennifer Doudna, co-inventor of CRISPR-Cas9, has long expressed caution about AI’s ability to generate truly novel scientific ideas. But, her lab just delivered a compelling counterpoint.

Using an inverse protein-folding AI model (inspired by Meta’s work), researchers redesigned TnpB — a compact, RNA-guided nuclease considered an evolutionary ancestor of CRISPR systems. They protected critical functional amino acids and allowed AI to optimise the rest for proper folding.

Key results: Nearly 2,000 designs were generated and tested. Roughly one in four showed DNA-cutting activity in bacteria, which is good, but the best performers also worked in plant and human cells. One top variant shared only 77% sequence identity with the natural enzyme - nearly a quarter of amino acids changed - yet retained functionality.

Why it matters for pharma: Smaller size improves delivery potential compared to larger CRISPR tools. The ability to design editors with tailored properties (specificity, activity, temperature stability) could accelerate development of gene therapies for a wide range of diseases. This shifts the field from evolutionary tinkering toward on-demand engineering.

Source: Fierce Biotech coverage

2. Precigenetics’ Cell Cinema: Watching Living Cells in Real Time

Traditional cell analysis often requires killing the sample. Startup Precigenetics is changing that with what they call “Cell Cinema,” a system that repeatedly observes the same living cell using laser light to read molecular bond responses - creating rich 3D chemical fingerprints of proteins, lipids, and metabolites without labels or destruction.

A microfluidic setup maintains cell viability while generating gigabytes of data per observation. Software turns these into dynamic trajectories showing how cells respond to treatments.

Early findings: Melanoma experiments suggest pre-treatment chemical profiles could predict drug resistance, and patterns of ferroptosis were detectable hours before standard assays.

Why it matters for pharma: This could revolutionise preclinical screening, mechanism-of-action studies, and resistance research. Imagine earlier, more nuanced readouts on drug efficacy and toxicity at the single-cell level. Caveats remain - the full paper is not yet public, and independent validation is pending - but the concept points to a less destructive, higher-resolution future for cell-based assays.

3. Chai Discovery Raises $400M at $3.8B Valuation with Lilly, Pfizer & Novartis On Board

Two years ago, Chai barely existed. Now it has closed a massive Series C and is working directly with top pharma players on its generative AI platform for molecular and antibody design.

Chai-3 reportedly doubles the success rate of its predecessor, delivering antibodies with up to 100× tighter binding in some cases, and achieves drug-comparable affinities for roughly half of targets - often after testing only small batches of designs.

Why it matters for pharma: This is validation at scale. Lilly, Pfizer, and Novartis are not just investors - they are active users integrating Chai’s tools into their discovery workflows. The model (provide the AI engine + custom training on proprietary data, without competing in the pipeline) appears pharma-friendly. It underscores a shift where computational design reduces the need for massive physical screening libraries and speeds up hits on difficult targets.

Source: Fierce Biotech

4. Lila Sciences: Toward AI Science Factories and “Local Spikes of Superintelligence”

Lila is building “AI Science Factories” - integrated systems connecting large models to physical labs (robots, instruments, and human operators). The AI proposes experiments, orchestrates execution, analyses results, and iterates.

In a recent Latent Space podcast, leaders described dramatic productivity gains (e.g., tiny teams advancing CAR-T programmes to non-human primate data in months) and emerging “local spikes” of superintelligent behaviour in specific scientific domains.

Why it matters for pharma: If the claims hold (and these are their claims…), AI-orchestrated labs could compress years of R&D into months at a fraction of the cost. Their “virtual startup” model lets partners outsource entire programmes. This represents the next evolution: not just AI for design, but AI closing the loop on physical experimentation. Public evidence is still emerging, but the vision aligns with broader industry pushes toward automated, data-rich discovery platforms.

The Bigger Picture

These updates reinforce a clear trend: AI is maturing from experimental aid to core infrastructure in pharma innovation. Whether it is redesigning molecular tools, generating higher-quality data from living systems, powering antibody pipelines with Big Pharma backing, or automating entire labs, the pace is accelerating.

Challenges remain - validation, reproducibility, integration into regulated workflows, and separating hype from deliverable value - but the momentum is undeniable. There’s lots to be across when you start researching the ‘state of the science’ in AI in pharma, but there’s not a day going by where you’d not want to be watching the news (or X)…

Further Reading & Sources

• Doudna lab work: Fierce Biotech article and recent Science paper coverage.

• Chai Discovery funding: Fierce Biotech

• Original roundup thread by Peter Ottsjö (BAIO): X / Twitter thread

• Lila Sciences insights: Recent Latent Space podcast interview with Andy Beam and Rafa Gómez-Bombarelli.

It’s a beautiful, mind-bending, fun idea about probability. It’s also, as recent mathematics has shown, profoundly misleading in any universe we actually inhabit.

The Maths Bites Back

A 2024 study (https://www.sciencedirect.com/science/article/pii/S2773186324001014) by mathematicians Stephen Woodcock and Jay Falletta put hard numbers on the fantasy. Even if every one of Earth’s roughly 200,000 chimpanzees typed one key per second on a 30-key keyboard for the entire remaining lifespan of the universe (the point at which heat death sets in), they would almost certainly never produce Shakespeare’s collected works (nearly 885,000 words of poetry, drama, and insight).

A single chimp, in those settings, has only a ~5% chance of randomly typing even the word “bananas” in its lifetime. The odds of producing even a short coherent sentence collapse into absurdity. The full canon would require impossible timescales.

The theorem survives as elegant mathematics, but as a practical intuition about creativity or discovery, it fails. Pure randomness at biological scales and finite time simply doesn’t cut it.

The Real Mistake Isn’t the Monkeys, It’s the Search

Here’s where the story gets interesting, and where it connects directly to how real progress happens in science, technology, and pharmaceutical innovation.

It’s tempting to say we’ve simply swapped in better monkeys: bigger, faster, digital ones. But that framing quietly repeats the same error the maths just challenged. Computers, and the AI systems running on them, aren’t chimpanzees at all, however fast. A chimp at a keyboard carries no information from one keystroke to the next; every letter is a fresh coin flip, which is exactly why the odds never improve, no matter how long it runs. An AI model does the opposite. Each token, each proposed molecule, each note in a melody is chosen *conditional* on everything that came before it and everything the model has absorbed about how language, chemistry, or music tends to behave. That conditioning is what collapses an astronomical search space down to a tractable one - not superior typing speed, but the elimination of almost every option that a blind search would have wasted time on.

That distinction matters more than it sounds. A model trained on the accumulated structure of English, Elizabethan drama, or protein-ligand binding isn’t sampling uniformly from the space of possible outputs - it’s sampling from a probability distribution shaped by everything humanity already knows about what tends to work. Large language models can now produce a decent, certainly coherent Shakespearean sonnet, a passable pastiche of a scene from *Hamlet*, or a plausible new plot structure within minutes. Not through brute-force randomness, but by having internalised the deep structures of form, psychology, and language, and then navigating that space with purpose. You can’t do it because you can’t internalize all of that - you can write something that’s passable, but you miss some of the things that matter.

Scale this with parallel compute, active learning, and rapid iterative feedback, and you move from “impossible by chance” to “routinely achievable with directed effort.”

The monkeys were never the point. The search strategy is.

The Real Lesson for Discovery

This isn’t just a thought experiment about literature. It’s a profound analogy for how innovation actually works - especially in fields like drug discovery, where the search space is astronomically larger than all of Shakespeare’s works combined.

Chemical space contains an estimated 10⁶⁰ or more drug-like molecules. Random screening of compounds is monkey-typing at planetary scale: slow, expensive, and mostly fruitless for complex targets.

The asymmetric advantage comes from replacing random search with intelligent, data-driven search: machine learning models that predict molecular properties before a single compound is synthesised, generative chemistry platforms that propose new molecular scaffolds optimised for binding, safety, and synthesisability, and closed-loop systems that design, simulate, test with real assay data, and learn from the result - active learning and Bayesian optimisation doing the work of deciding what to try next, rather than reinforcement learning tuned to human preference, which is a different tool built for a different job.

This isn’t hypothetical. Insilico Medicine’s rentosertib - a TNIK inhibitor for idiopathic pulmonary fibrosis, a disease with no treatments that reverse its progression - had its target identified by an AI platform trawling multi-omics data, and its molecular structure designed by a generative chemistry engine. It entered Phase III trials in July 2026. That’s the pipeline this post is actually about: not a thought experiment, a drug in human trials right now.

What once required screening millions of compounds over years can now be guided toward promising regions of chemical space in days or weeks. We’re not waiting for infinite random trials. We’re building better searchers.

This is the core of asymmetric learning: the organisations and teams that develop superior ways of navigating vast possibility spaces - through better data, better models, better feedback loops - pull ahead dramatically. The difference between random exploration and directed intelligence isn’t incremental. It’s transformative.

The Catch: Nobody Knows (Knew) What Shakespeare Looks Like in Advance

There’s a limit to the monkey analogy worth naming honestly, because it’s the difference that actually matters for anyone doing this work.

The monkeys are searching for a known, hard target. We have the complete works of Shakespeare already; success is an exact match against a fixed string. Drug discovery doesn’t work that way, and neither does genuine creative work. It’s a question no-one has the answer to - yet. Nobody has the answer key for the molecule that will safely treat a disease with no existing treatment, or for the novel that hasn’t been written yet. The objective itself is uncertain, contested, and often redefined mid-search, as new safety data or new critical judgement reshapes what “success” even means.

That’s a harder problem than typing towards Shakespeare, in one sense - and an easier one in another. You can’t verify progress against a known answer, but you also aren’t constrained to reproduce something that already exists. The real work of directed search isn’t just navigating towards a target efficiently. It’s deciding, continuously, what the target should be.

Beyond Parroting: Toward Genuine Novelty

Critics will rightly note that current AI systems are sophisticated remixers and predictors rather than true originators *ex nihilo*. They still benefit enormously from human curation, taste, and judgement.

But it’s worth asking how much of a contrast that really is with Shakespeare himself. He didn’t emerge from a probability distribution, but he wasn’t working from nothing either. He was a highly trained pattern-matcher, with Plutarch, Holinshed, Italian novellas in his reading list (unlike mine, he put them on his bookshelves once read), and the conventions of the sonnet and the five-act structure, iterating within tight formal constraints and a specific culture. Directed search, in other words, isn’t unique to machines - it’s arguably a reasonable description of how human creativity has always worked, at a much smaller and slower scale. It’s writing a song when someone tells you the key, the time signature, and the length and style. The interesting frontier isn’t “AI versus human insight” as two opposed categories. It’s how the two forms of directed search compose: AI proposing, humans steering and selecting on taste and judgement that no model yet has, AI iterating again.

In pharma, this already looks like AI proposing candidates that medicinal chemists would never have reached for, followed by expert refinement and experimental validation. Each cycle narrows the gap between what’s proposed and what actually works.

The Cost of Getting It Wrong Quietly

There’s a failure mode worth naming, because it’s asymmetric in a way that’s easy to miss.

A chimpanzee’s garbled text is obviously garbled - nobody mistakes even that banana example, “xjqpz banana zzzqx”, for a sonnet. An incorrect answer from a directed search doesn’t announce itself in the same way. A molecule can look right on paper - clean binding predictions, plausible synthesis route - and still fail in a way that only shows up in a clinical trial. A generated passage can scan perfectly and say nothing. The risk of intelligent search isn’t that it fails more often than random search; it’s that when it fails, it can fail *convincingly*, in a form that passes casual inspection precisely because it was optimised to look right.

That’s a genuinely asymmetric cost, and it’s the reason validation loops - real assay data, real experimental feedback, real human judgement in the loop - aren’t a bolt-on to directed search. They’re what keeps the asymmetry working in your favour instead of against you.

What This Means Practically

If you work in R&D, strategy, or innovation:
- Stop thinking in terms of “more shots on goal” through brute-force randomness.
- Start thinking in terms of better aim through superior models and feedback, and be precise about which technique is actually doing the aiming - active learning and Bayesian optimisation for scientific search, not tools built for aligning conversational tone.
- Invest in the infrastructure that turns data into directed intelligence: high-quality datasets, simulation capabilities, rapid experimental loops, and human-AI collaboration workflows.
- Build in validation that catches confidently-wrong answers, not just obviously-wrong ones - the failure mode that matters most is the plausible one.
- Measure not just activity (compounds screened, models trained) but asymmetric progress - how much faster or more effectively you’re navigating the space that matters.

The infinite monkey theorem was always a story about limits. The real story today is about transcending those limits through intelligence and design - not by building a better monkey, but by leaving the whole metaphor behind.

We don’t need every chimpanzee on Earth typing forever. We need better-directed search, pointed at problems worth solving, with humans still deciding what “worth solving” means.

That’s how the next breakthroughs - in literature, in science, in medicine - will actually arrive.

If you could play chess against an opponent, but use an AI to guide your moves against an opponent who wasn’t using help, how do you think you’d do? In pharma, you can.

When Deep Blue beat Kasparov in 1997, it felt like a threshold. It was - and we were never going back. Engines like Stockfish didn’t catch up with top humans; they left them behind. Today the frontier has moved into competitions like TCEC, where machine plays machine and progress comes from relentless iteration, testing, and refinement.

Once the best systems surpassed the best people, chess stopped being a story about humans playing better. It became a story about different kinds of learning systems competing with one another. The tech companies worked out that the way to win isn’t just better chips - it’s how they learn. And they do want to win.

I think drug discovery is moving, unevenly but unmistakably, towards a similar transition. Not because biology is as tractable as chess (it isn’t), not because AI will replace scientists (it won’t), and not because pharma innovation reduces to search plus compute (it doesn’t). But because the next wave of advantage in R&D won’t come from playing the old game slightly better. It will come from designing systems that learn asymmetrically: faster, broader, more honestly, and with better feedback than traditional organisations can manage.

From symmetric improvement to asymmetric advantage

Traditional chess mastery was symmetric - study openings, annotate games, train memory, sharpen intuition. Hard, disciplined, bounded by human constraints: limited memory, cognitive bias, emotional attachment to prior ideas, the cost of exploring only a small slice of the search space.

AlphaZero and LCZero changed the terms. They didn’t replicate grandmaster expertise more efficiently. They learned under a different regime: vast self-generated experience, rapid feedback, continual reweighting, freedom from sunk costs. The machine advantage didn’t come from doing grandmaster things faster. It came from learning under a different regime.

And that is exactly the question pharma now faces.

The enduring symmetry of pharmaceutical R&D

Much of pharma R&D still follows a symmetric logic. The pipeline is linear. The bets are large. Decisions get framed around confidence, precedent, and the appearance of de-risking rather than around maximising learning under uncertainty. Programmes gather momentum. Weak signals get rationalised. By the time an asset reaches a decisive stage, enormous money, time, and internal narrative have accumulated behind it.

Caution is rational here - drug development isn’t a game. But caution easily hardens into symmetry: doing familiar things in familiar sequences, optimising locally, making fewer larger bets and expecting late-stage success to justify earlier uncertainty. The outcomes are familiar too: high attrition, late learning, expensive surprises. That’s not a failure of intelligence. It’s a failure of learning architecture.

What chess engines reveal

The best engines search more broadly, update more quickly, receive cleaner feedback, and operate at a tempo no human learner can match. Translate it carefully into drug discovery and four things follow.

Search. Don’t just mimic the expert’s intuition — surround it with broader, disciplined exploration. Generative models, active learning that picks the next compound by expected information gain, structure-based modelling on the back of AlphaFold. The value isn’t one more attractive compound; it’s a system that explores dozens of serious options early and learns proprietary insight from the pattern of successes and failures. This is why I keep returning to “explore 100 options per asset early” - not because all 100 are equally good, but because the team that examined five can’t catch up.

Feedback. Every experiment should teach more than one thing. Self-play systems are built around dense feedback - every game is data, every iteration updates the policy. Pharma still tends to test a compound, get a readout, answer one question, move on. An asymmetric model treats each experiment as a system-level update: assay results refine the design model, failed compounds become informative examples, Phase I and Phase II become information engines rather than gates to be crossed.

Infrastructure. Compute alone isn’t the answer, but it’s not optional. If your biological data is fragmented, poorly standardised, ontologically unstable, models trained on it will reflect those weaknesses. Poor data architecture forces organisations back into symmetric habits because the system can’t support richer iteration. Robust infrastructure is increasingly the substrate through which scientific organisations learn - and learning speed, not just operational efficiency, becomes the strategic asset.

Competition. TCEC rewards structured competition between strong alternatives: clones get penalised, weaknesses get exposed. Pharma talks about innovation but internally suppresses exactly this kind of disciplined competition. A lead asset acquires champions, alternative hypotheses become politically awkward, weakening evidence gets reinterpreted to preserve coherence. Take the analogy seriously and you’d organise differently: multiple hypotheses competing under shared criteria, aggressive benchmarking, cognitive diversity as protection against conceptual inbreeding.

The crucial difference: biology has no clean reward function

This is where the chess analogy is most fragile and most important. Chess is uniquely hospitable to optimisation - fixed rules, visible board, unambiguous objective. Drug discovery is nothing like that. A molecule isn’t “good” in any single sense: potency, selectivity, toxicity, PK, developability, patient benefit, reimbursement, real-world outcomes all matter, and they can diverge. The reward function is plural, contested, and deeply human.

That’s why alignment matters so much in medicine. As Brian Christian argues, systems optimise what we specify, not what we vaguely intend. Self-improving systems become powerful only when their goals, constraints, proxies, and feedback loops are designed with unusual care - as much an institutional and philosophical challenge as a technical one. Human judgement remains central, not as romantic defence against machines, but as a necessary part of setting objectives, spotting bad proxies, and deciding which trade-offs are acceptable.

Why symmetric AI adoption will disappoint

The biggest near-term risk isn’t that pharma ignores AI. It’s that it adopts AI symmetrically: bolting generative tools onto traditional funnels, using prediction to justify prior beliefs, demanding immediate ROI, optimising for cosmetic speed rather than epistemic quality. That’s like giving grandmasters slightly better opening databases while insisting the real contest is unchanged.

The deeper opportunity is to redesign the system itself: more parallel exploration, richer feedback, earlier error detection, more honest portfolio decisions, teams rewarded for updating rather than defending. That’s not a tool problem. It’s an operating-model problem.

What a TCEC-style R&D organisation might actually do

Take the analogy seriously at organisational level and seven design principles follow:

1. Run more meaningful parallel bets early - not indiscriminate proliferation, but disciplined breadth across compounds, mechanisms, biomarkers, and disease framings where appropriate.

2. Optimise for learning velocity, not just asset progression - a programme generating fast, decision-relevant knowledge may be more valuable than one that advances smoothly while concealing structural weakness.

3. Close the loop between models and experiments - predictions alter experimental design, outcomes retrain models, the cycle is continuous not episodic.

4. Reward rapid disconfirmation - in many organisations being wrong early is career-limiting while being wrong late is socially survivable because everyone was wrong together. That’s a disastrous incentive structure.

5. Benchmark aggressively - compare models, compounds, assay strategies, translational assumptions, and portfolio narratives under explicit criteria. Strong systems improve through challenge.

6. Invest in conceptual infrastructure, not just technical tooling - ontologies, data standards, target definitions, endpoint logic. Sloppy conceptual architecture quietly destroys learning efficiency.

7. Keep humans in the loop in the right role - framing the search well, specifying objectives wisely, spotting confounded proxies, deciding which uncertainties are worth paying to resolve. That’s not a diminished role. It’s a higher one.

The companies worth watching

A few useful archetypes are emerging. Recursion is building a computation-first discovery platform with large-scale biological data and iterative wet-lab feedback loops. Isomorphic Labs is pursuing the convergence of frontier AI and biology from a more fundamental-science angle. Insilico Medicine and Schrödinger represent different bets on how modelling, design, and simulation can reshape parts of discovery. None is proof the problem is solved - biology is too difficult, drug development too path-dependent, for that confidence. What they offer are experiments in alternative learning architectures. The interesting question isn’t which company has the best model; it’s which organisation is building the best system for generating, testing, and updating knowledge under real biological uncertainty.

Towards an asymmetric future

The human era in chess didn’t end because better humans arrived. It ended because superior learning systems did. Drug discovery won’t mirror that story exactly - biology is harder than chess, medicine answers to patient welfare, the objective function is messier. But the direction of travel is unmistakable. Imagine a world in which classic grandmasters could consult AI before their moves. That’s an imaginable world for us in pharma.

The winners won’t be those with the largest pipelines or the noisiest AI rhetoric. They’ll be organisations that build systems capable of exploring more broadly, learning more quickly, updating more honestly, and aligning their optimisation with real human and scientific value. They’ll treat uncertainty not as an embarrassment to be hidden until late-stage trials, but as terrain to be navigated deliberately. They’ll use AI not as a substitute for judgement, but as an amplifier of asymmetric exploration.

In chess, the strongest engines didn’t win by becoming more human. In drug discovery, the same principle applies. The question isn’t whether AI matters. It already does. The question is whether we’ll use it to automate yesterday’s pipeline - or to build better engines for learning.

I’m aware that it’s 30 years since humans ceded chess to the machines, but I’m still a fan of the humans’ quotes. I made my own from Garry Kasparov’s quotes:

“The essence of strategy isn’t finding the best move - it’s correctly evaluating which of your assumptions are wrong before your opponent forces you to find out the hard way.”

“It is still impressive how many poetical blunders derive from ‘obvious’ assumptions.” Garry Kasparov, How Life Imitates Chess

“It is so important to question success as vigorously as you question failure.” Garry Kasparov, How Life Imitates Chess

For most of modern history, chess has served as one of our favourite metaphors for human intelligence. It condenses so much of what we admire in the mind: strategic depth, pattern recognition, memory, intuition, discipline, foresight. It is not something I ever had the patience to master myself, but I have always been drawn to it as an example of what serious learning looks like.

Then the machines took over.

When IBM’s Deep Blue defeated Garry Kasparov in 1997, it felt like a threshold moment - not only in computing, but in culture. A machine had beaten the world champion at a game long treated as a citadel of human thought. Yet in retrospect, that was merely an early milestone - clearly it was a threshold, and we were never going to go back to a time when we’d beat them. What followed was perhaps more interesting. Engines such as Stockfish did not simply catch up with top human players; they left them behind entirely. Today, the strongest humans do not meaningfully compete with the strongest engines. The frontier has moved on, into competitions like TCEC, where machine plays machine and progress comes from relentless iteration, testing, and refinement. Just consider how much, for humans, this must resemble ants watching us and trying to figure us out…

That shift matters. Once the best systems surpassed the best people, chess ceased to be primarily a story about humans playing better. For me, it has become a story about different kinds of learning systems competing with one another - the tech companies have figured out that the way to win isn’t just better chips, it’s about how they learn.

I think drug discovery is moving, unevenly but unmistakably, towards a similar transition.

Not because biology is as tractable as chess - it is not. Not because AI will replace scientists - it will not. And not because pharmaceutical innovation can be reduced to search plus compute. It cannot. But because the next wave of advantage in R&D is unlikely to come from doing the old game slightly better. It will come from designing systems that learn more asymmetrically: faster, broader, more honestly, and with better feedback than traditional organisations can manage.

That is the deeper analogy. I hope you’ll forgive me for deepening further. As I say towards the end, imagine grandmaster chess players who are able to use laptops and AI in their games. Kind of like the (in reality, underwhelming) Enhanced Games, but for chess. Unlikely to happen, that analogy is definitely a reality for pharma.

From symmetric improvement to asymmetric advantage

My own bias, as readers will know, is to think about progress through the lens of asymmetric learning: situations in which some approaches generate unevenly large gains in understanding relative to the time, effort, or capital invested. This is not simply about being smarter or working harder. It is about structuring the learning process so that each cycle yields more signal, more insight, and more strategic leverage than a conventional one.

Traditional mastery in chess was largely symmetric. You studied openings, annotated games, learned tactical motifs, built endgame knowledge, trained memory, sharpened intuition, and improved through experience. None of this was trivial; indeed, it represented, for many, one of the highest forms of disciplined human learning, and mastery. But it was bounded by recognisably human constraints: limited memory, limited time, cognitive bias, emotional attachment to prior ideas, and the sheer cost of exploring only a small fraction of the relevant search space.

Computer chess changed the terms of the contest. Early engines relied heavily on brute-force search, exploring many possible lines more quickly than humans could. But later systems became more interesting still. AlphaZero, for example, demonstrated the power of reinforcement learning and self-play at extraordinary scale, while LCZero carried related ideas into an open chess-engine ecosystem. These systems did not merely replicate human expertise more efficiently. They learned in a fundamentally different way: through vast self-generated experience, rapid feedback, continual reweighting of strategies, and freedom from many of the heuristics and sunk costs that constrain human thought.

This is the heart of the matter. The machine advantage in chess did not come simply from doing grandmaster things faster. It came from learning under a different regime.

And that is exactly the question pharma now faces.

The enduring symmetry of pharmaceutical R&D

Despite periodic talk of transformation, much of pharmaceutical R&D still follows a deeply symmetric logic.

The pipeline is linear (often symmetrical with competitors). The bets are large. The process is slow. Decisions are often framed around confidence, precedent, and the appearance of de-risking, rather than around maximising learning under uncertainty. Expert judgement is indispensable, but often overburdened. Historical data are used heavily, but not always in ways that genuinely update belief. Programmes gather institutional momentum. Weak signals are rationalised. By the time a major asset reaches a decisive stage, an enormous amount of money, time, and internal narrative has already accumulated behind it.

This structure is understandable. Drug development is not a game. The stakes are ethical, clinical, regulatory, and financial. False confidence can harm patients. Noise is endemic. Ground truth is expensive. Biology is not fully observable. In such an environment, caution is rational. But caution can easily harden into symmetry.

By symmetry here, I mean a mode of learning in which organisations mostly do familiar things in familiar sequences, with only incremental improvements in speed or sophistication. They optimise locally. They move stage by stage. They seek cleaner narratives than the underlying science deserves. They make fewer, larger bets and expect late-stage success to justify earlier uncertainty.

The outcomes are familiar too: high attrition, late learning, expensive surprises, and productivity gains that remain modest relative to the scale of investment.

That is not a failure of intelligence. It is a failure of learning architecture.

What chess engines reveal about the structure of advantage

The best chess engines are not impressive simply because they win.

They search more broadly.
They update more quickly.
They test more possibilities.
They receive cleaner feedback.
They are less attached to human priors.
They improve through structured competition.
And they operate at a tempo no human learner can match.

That is a useful template for thinking about drug discovery, for as long as we translate it carefully.

The aim is not to imagine a magical system that autonomously discovers medicines while people stand back and wait for them to hit the meat grinder of phase II/ III. The aim is to ask how an R&D organisation might gain disproportionate advantage by redesigning its loops of hypothesis generation, experimentation, feedback, and decision-making.

Search: from a few polished ideas to broad structured exploration

A good medicinal chemist carries a remarkable internal model: scaffold intuition, SAR pattern recognition, synthetic tractability, practical know-how, awareness of historical precedents, and the kind of tacit judgement that does not reduce neatly to rules. That expertise remains essential. But it is still bounded, by their own version of the ‘books of plays’ that limit the grandmasters.

The opportunity with modern AI and computation is not merely to mimic that expertise. It is to surround it with a much broader and more disciplined search process.

That can take several forms: virtual screening across vast chemical libraries, generative models proposing novel molecules, active-learning systems choosing which compounds to make next based not just on predicted activity but on expected information gain, and structure-based modelling informed by advances such as AlphaFold and related tools in protein structure prediction.

AlphaFold matters here not because it solved biology - it plainly did not - but because it showed how a sufficiently powerful learning system, trained on the right representations and objectives, could dramatically shift one part of the scientific workflow. It changed expectations about what kinds of inference were computationally plausible. It also reminded us that once a bottleneck begins to move, adjacent bottlenecks suddenly matter more.

The same is true in medicinal chemistry. Once it becomes cheap to generate many more plausible candidates in silico, the question becomes: how do you choose among them intelligently, learn from them efficiently, and avoid simply moving the bottleneck downstream?

That is where asymmetric learning starts to matter. The real value is not that one model produces one more attractive compound. It is that a system can explore dozens or hundreds of serious options early, and learn from the pattern of successes and failures in ways that become hard for competitors to replicate. This is one reason I keep returning to the idea of exploring 100 options per asset early on: not because all 100 are equally good, but because the process of exploring them can generate proprietary insight unavailable to a team that only ever examined five.

In that sense, early discovery should be understood less as funnel management and more as search under uncertainty with compounding informational returns.

Feedback: every experiment should teach more than one thing

One of the profound advantages of self-play systems such as AlphaZero is that they are built around dense feedback. Every game is data. Every iteration updates the policy. The system does not merely test whether one move was good; it refines its entire approach through cumulative exposure to structured outcomes.

By contrast, many experiments in pharma still function too narrowly. A compound is tested. A readout comes back. A specific question is answered. Then the organisation moves on. Valuable information is produced, but too often it is not fully captured as part of a broader learning loop.

A more asymmetric model would treat each experiment as a contributor to a system-level update.

That means assay results should refine the design model, not simply rank compounds. Failed compounds should not merely be discarded; they should become informative examples. Pharmacokinetic surprises should alter the search process. Translational mismatches should update not only confidence in one programme, but the architecture of how programmes are selected and advanced.

This is particularly important in clinical development. Too many firms still describe early clinical stages as hurdles to be crossed en route to pivotal trials. But Phase I and Phase II can be understood differently: as high-value information engines. They can reveal mechanism validity, dose-response, biomarker utility, patient heterogeneity, endpoint quality, and whether the story the organisation has been telling itself actually survives contact with patients.

That shift in framing matters. If early trials are treated only as gates, organisations maximise the chance of passing through them. If they are treated as learning systems, organisations maximise the value of what they learn from them - even when the answer is uncomfortable.

The latter is much closer to an asymmetric advantage.

Infrastructure: compute is not enough, but it is not optional either

Chess engines benefit from superior compute, but their strength is not reducible to hardware. Compute matters because it supports search, iteration, and evaluation inside a well-designed system.

Pharma has an analogous challenge. Too often, data platforms, interoperability, automation, multimodal integration, and model infrastructure are treated as support functions. They are funded defensively and discussed operationally. Yet these are increasingly the substrate through which scientific organisations learn.

If your biological data are fragmented, poorly standardised, ontologically unstable, and slow to integrate, then no amount of excitement about AI will rescue you. Models trained on noisy, inconsistently labelled, weakly connected data will reflect those weaknesses. This is one reason the ontology and noise problem in molecular machine learning matters so much. The issue is not merely technical hygiene. It is that poor conceptual and data architecture forces organisations back into symmetric habits, because the system cannot support richer forms of iteration.

Conversely, firms that build robust data infrastructure, tighter model-experiment loops, and better ways of integrating real-world, preclinical, and mechanistic evidence may gain more than operational efficiency. They may gain learning speed.

That, in many settings, is the real strategic asset.

Competition: the underrated engine of better science

One feature of elite computer-chess competitions that deserves more attention is that they reward structured competition between strong alternatives. TCEC is not just a (kind of) spectacle. It is a system in which high-quality opponents, strong evaluation, and iterative updates produce rapid improvement. Clones are penalised. Innovation matters. Weaknesses are exposed ruthlessly.

Pharma talks often enough about innovation, but internally many organisations suppress precisely the kind of disciplined competition that would improve their judgement.

A programme gains momentum. A lead asset acquires champions. Alternative hypotheses become politically awkward. Internal debate narrows. Weakening evidence is reinterpreted to preserve coherence. A mechanistic story becomes fashionable. (Look - I’m a positioning guy, and I know how even the story is usually taken via the easiest route - there’s rarely competition between stories, and teams rarely look deep enough to find the killer story… Mechanism of effect, vs mechanism of action, is rarely even examined…)

This is not always conscious. Often it is simply what large institutions do when incentives reward narrative stability over honest updating.

But if you take the engine analogy seriously, you would organise differently. You would allow multiple hypotheses to compete under shared criteria. You would benchmark models and programmes against one another explicitly. You would encourage serious challenge from outside the dominant coalition. You would value cognitive diversity not as a slogan, but as protection against conceptual inbreeding.

That last point matters more than it may appear. Some of the most interesting frontier AI groups have hired philosophers and other non-traditional thinkers not out of eccentricity, but because problem-framing, alignment, and abstraction are strategic capabilities. Cross-domain thinking often looks ornamental to symmetric organisations, right up until it becomes decisive. I’ve examined the power of the word ‘if’, which has informed my new company, here…

Drug discovery does not need performative interdisciplinarity. It does need teams capable of seeing the system from more than one angle.

The crucial difference: biology has no clean reward function

This is where the analogy with chess becomes most fragile, and most important.

Chess is almost uniquely hospitable to optimisation. The rules are fixed. The board is visible - your opponent can literally see every move you make. The objective is explicit. Win, lose, or draw. The feedback may be complex, but the end-state is unambiguous.

Drug discovery is not like that at all. A molecule is not “good” in any single sense. Potency matters, but so do selectivity, toxicity, pharmacokinetics, developability, manufacturability, route of administration, patient adherence, clinical differentiation, regulatory strategy, cost, reimbursement, and ultimately real patient benefit. Even then, biology may surprise you (it usually does, if we take attrition as a guide). The same mechanism can behave differently across patient groups. A trial endpoint can flatter a weak therapy or obscure a useful one. Safety signals may appear late. Commercial and clinical value may diverge.

In other words, the reward function is not just complex. It is plural, contested, and deeply human.

That is why the question of alignment matters so much in medicine. Brian Christian's The Alignment Problem is helpful here, not because it gives a neat recipe, but because it insists on a distinction that is easy to forget: systems optimise what we specify, not what we vaguely intend. In chess, specifying the objective is easy. In medicine, it is profoundly difficult.

The lesson, then, is not that pharma should copy AlphaZero literally. It is that self-improving systems become powerful only when their goals, constraints, proxies, and feedback loops are designed with unusual care. That is as much an institutional and philosophical challenge as a technical one.

It is also why human judgement remains central - not as a romantic defence against machines, but as a necessary part of setting objectives, detecting bad proxies, and deciding which trade-offs are ethically and scientifically acceptable.

Why symmetric AI adoption will disappoint

The greatest near-term risk is not that pharma ignores AI. It is that it adopts AI symmetrically.

By that I mean using models to accelerate existing steps without changing the underlying structure of learning. Bolting generative tools onto traditional discovery funnels. Using prediction systems mainly to justify prior beliefs. Demanding immediate ROI from each isolated application. Treating models as authoritative rather than exploratory. Optimising for cosmetic speed rather than epistemic quality.

Done this way, AI will produce some value, but not a step-change. The result will be incremental acceleration inside an architecture still dominated by late learning, fragile narratives, and concentrated bets.

That would be like responding to chess engines by giving grandmasters slightly better opening databases while insisting the real contest remained unchanged.

The deeper opportunity is to redesign the system itself: more parallel exploration, richer feedback loops, earlier and cheaper error detection, more honest portfolio decisions, stronger data architecture, and teams rewarded for updating rather than defending.

That is not a tool problem. It is an operating-model problem.

What a TCEC-style R&D organisation might actually do

If you take this analogy seriously at organisational level, several design principles follow.

First, run more meaningful parallel bets early. Not indiscriminate proliferation, but disciplined breadth across compounds, mechanisms, biomarkers, and even disease framings where appropriate.

Second, optimise for learning velocity as well as asset progression. A programme that generates fast, decision-relevant knowledge may be more valuable than one that advances smoothly while concealing structural weakness.

Third, close the loop between models and experiments. Predictions should alter experimental design. Experimental outcomes should retrain models. The cycle should be continuous, not episodic.

Fourth, reward rapid disconfirmation. In many organisations, being wrong early is career-limiting, while being wrong late is socially survivable because everyone was wrong together. That is a disastrous incentive structure. An asymmetric organisation treats early error detection as an asset.

Fifth, benchmark aggressively. Compare models, compounds, assay strategies, translational assumptions, and portfolio narratives under explicit criteria. Strong systems improve through challenge.

Sixth, invest in conceptual infrastructure, not just technical tooling. Ontologies, data standards, target definitions, endpoint logic, and shared conceptual clarity matter enormously. Sloppy conceptual architecture quietly destroys learning efficiency.

Seventh, keep humans in the loop in the right role. The value of scientists and clinicians increasingly lies not in manually searching every possibility, but in framing the search well, specifying objectives wisely, spotting confounded proxies, interpreting weak signals, and deciding which uncertainties are worth paying to resolve.

That is not a diminished role. It is a higher one.

The companies worth watching

I would be careful here not to lapse into hype, because this space is full of narratives with no evidence. But there are at least a few useful company archetypes to watch.

Some firms are trying to build computation-first discovery platforms with large-scale biological data and iterative wet-lab feedback loops - Recursion is one prominent example. Others, such as Isomorphic Labs, are pursuing the convergence of frontier AI and biology from a more fundamental-science angle. Still others, including Insilico Medicine and Schrodinger, represent different bets on how modelling, design, simulation, and iteration can reshape parts of the discovery process.

I would not cite any of these as definitive proof that the problem is solved. Biology is too difficult, and drug development too path-dependent, for that kind of confidence.

What they do offer, though, are experiments in alternative learning architectures. That is the more interesting lens. The question is not simply which company has the best model. It is which organisation is building the best system for generating, testing, and updating knowledge under real biological uncertainty.

Towards an asymmetric future in drug discovery

The human era in chess did not end because better humans arrived. It ended because superior learning systems did.

Drug discovery will not mirror that story exactly. The analogy is too imperfect for that. Biology is harder than chess. Medicine answers to patient welfare, not game outcomes. The objective function is messier, the feedback slower, the search space more treacherous.

And yet the direction of travel is unmistakable. Imagine a world in which classic grandmasters could consult AI before their moves. That is an imaginable world for us in pharma…

The winners in this next era are unlikely to be those who merely accumulate the largest pipelines, the noisiest AI rhetoric, or the most polished strategy decks. They will be the organisations that build systems capable of exploring more broadly, learning more quickly, updating more honestly, and aligning their optimisation with real human and scientific value.

They will treat uncertainty not as an embarrassment to be hidden until late-stage trials, but as a terrain to be navigated deliberately. They will use AI not as a substitute for judgement, but as an amplifier of asymmetric exploration. They will invest in feedback loops, conceptual clarity, structured competition, and the institutional capacity to change course when reality demands it.

In chess, the strongest engines did not win by becoming more human.

In drug discovery, the same principle applies. The question is not whether AI matters. It already does. The question is whether we will use it to automate yesterday’s pipeline - or to build better engines for learning.

That’s the pattern I’ve been calling a “collection of biotechs” - big pharma building out its pipeline not through internal discovery but by buying the companies that already did the ‘hard’ part. Kyle LaHucik at Endpoints News captured the mood neatly a few days ago: biopharma, in his words, just keeps buying, buying, buying. By his tally, the first half of 2026 produced 45 acquisitions worth a combined $125.5 billion - which looks likely to beat last year’s 61 deals for the full year, and July has already added four more. (Other trackers land on different totals depending on what counts as a “deal” - BioSpace has 52, PitchBook nearer 200 once you include smaller bolt-ons - so I’d treat any single figure as directional rather than gospel. The direction, at least, isn’t in dispute.)

The rest of the week’s haul

Vertex wasn’t dealmaking alone. Ipsen signed twice in a single week: up to $1.75 billion for Kartos Therapeutics and its Phase III myelofibrosis drug navtemadlin, followed almost immediately by roughly $800 million for Memo Therapeutics, whose lead antibody treats BK polyomavirus-associated nephropathy in kidney transplant patients - a nasty, underserved complication with no approved targeted treatment. Novartis added Myricx Bio, a UK antibody-drug conjugate specialist, for up to $1.5 billion - its third deal this year. And United Therapeutics picked up Thymmune Therapeutics for up to $300 million, a smaller bet on regenerative thymic cell therapy for transplant patients.

Different sizes, same logic. None of these companies needed to build acromegaly expertise, ADC chemistry or thymic biology from scratch. They bought it, fully formed, from the biotechs that spent years getting the science right.

Why buy rather than build

I’ve made the underlying case before: on a risk-adjusted basis, acquiring proven or near-proven science tends to beat internal discovery, because you’re paying for assets that have already cleared the parts of drug development where most failure happens. If you want the numbers behind that - success rates by phase, the Amgen-versus-GSK comparison, what a de-risked asset is actually worth - I laid it out in an earlier post, and this week’s deals are a fairly clean illustration of the argument in practice.

But there’s a catch I flagged in that piece and haven’t resolved yet, and it matters more with every deal like this one: buying the science is the easy part. Keeping what made that science good in the first place is the hard part.

What this does to valuations

There’s a second effect worth naming: this pace of dealmaking is quietly repricing risk further back in the pipeline than it used to be. A few years ago, the smart exit for a lot of biotechs was to partner off an asset after decent Phase II data and let a pharma partner carry the pivotal trial risk. Look at what’s just been paid for, though. Kartos was still Phase III, not yet approved, and still fetched $1.75 billion. Memo’s lead antibody is barely out of Phase II and Ipsen is putting up to $800 million behind it. Even Thymmune, earlier-stage and more speculative than either, found a buyer willing to commit up to $300 million. If that pattern holds, running your own Phase II - or pushing on into Phase III rather than out-licensing early - starts to look like a better-compensated bet than it did two years ago. You’re not de-risking the asset purely for someone else’s benefit; you’re capturing more of that de-risking in the price you’re eventually paid.

The obvious counter is survivorship bias: for every Kartos, there’s a Phase III that reads out badly and quietly disappears, and a week of headlines only ever shows you the winners. I’d also guess this compresses unevenly - a well-understood mechanism in a hot area like oncology or endocrinology probably gets rewarded for going longer; a genuinely novel modality with no read-through comparable asset probably doesn’t. But at the margin, with acquirers this hungry, holding an asset one stage longer than you used to is starting to look like the more rational play rather than the riskier one. Worth watching whether valuations for pre-Phase-III biotechs start drifting up even before data reads out, on the assumption that a buyer will be waiting either way.

The bit that actually worries me

Here’s what I actually believe, for what it’s worth: biotechs are good at drug development not despite being small, underfunded and slightly chaotic, but because of it. The founder who’s spent eighteen years on one disease, the team that’s small enough to change course on a Tuesday, the culture that tolerates the sort of expensive failure a public pharma company would never wave through committee - that’s not incidental to the science. It’s a big part of where the science comes from.

Scott Struthers spent nearly eighteen years building Crinetics around endocrine disease before selling to Vertex this week. That kind of focus is exactly the “collection” model is meant to acquire - and exactly what’s hardest to keep alive once a company the size of Vertex owns you. Big pharma is very good at commercialisation, regulatory strategy and capital allocation. It is not obviously good at leaving a newly-acquired team alone to keep being weird and stubborn about the thing it’s good at. (I do believe Lilly might be an outlier, as I covered in my interview with Verve’s Sek Kathiseran.)

None of this week’s deals are old enough to tell us anything yet - Vertex/Crinetics doesn’t even close until Q3. But it’s worth watching: does Crinetics keep some form of independent identity, or does it simply become another line on Vertex’s org chart within eighteen months? Same question for Ipsen’s two targets, bought within days of each other. My hunch, and it’s only a hunch so far, is that deals built to preserve some autonomy - separate sites, retained leadership, backloaded earn-outs like Ipsen structured for Memo - hold onto more of the value they paid for than deals that go straight to full integration. There was nothing sadder than watching Genentech get ‘Roche-d’ in the ‘olden’ days.

Where this leaves things

The collection strategy isn’t a bubble, and it isn’t going away - the incentives (patent cliffs, cash-rich balance sheets, a public market that rewards pipeline more than lab spend) are too strong. But “big pharma is buying a lot of biotechs” is the easy story. The harder, more interesting one is what survives the acquisition besides the molecule. I’ll be tracking that as these deals actually close, and the drugs launch, rather than just when they’re announced.

Sources: Endpoints News (Kyle LaHucik), company press releases, BioPharma Dive, BioSpace, and my earlier post on acquiring biotech innovation, linked above.

A model designs a novel molecule. The molecule enters Phase I. The trial clears its safety hurdle. The company raises money, the coverage declares another milestone, and everyone agrees that artificial intelligence is finally beginning to crack one of pharma’s hardest problems.

I think this is almost exactly backwards.

AI may well be improving drug discovery. In some areas, it clearly is. The problem is that the industry has spent the past decade optimising the part of drug development that was least bottlenecked - while the expensive, slow, patient-constrained part of the system remains largely unchanged.

That is the funnel fallacy.

The bull case says AI gives pharma more shots on goal.

But pharma’s deepest problem was never a shortage of shots. It was that almost nothing scores.

Or, more precisely: the value of more shots depends entirely on whether they help you learn faster before the expensive decisions arrive.

That is the distinction most of the AI drug discovery debate misses.

The win is not more molecules. The win is more asymmetric learning per dollar.

The numbers everyone likes to quote

As of early 2026, more than 173 AI-discovered drug programmes are in clinical development. That is the number that gets the press release. The more interesting number is 15 to 20 - the number of those programmes expected to reach Phase III this year.

The gap between those two numbers is the whole story.

To be clear, some of the early data is genuinely impressive. AI-discovered molecules appear to be clearing Phase I at somewhere around 80 to 90 percent, compared with a historical average of roughly 52 percent for traditionally developed compounds.

That is not a rounding error. It is a real improvement.

It also makes intuitive sense. Phase I is mostly, traditionally, a safety and tolerability screen. It asks whether a compound is obviously toxic, whether it behaves in the body in a broadly acceptable way, and whether it can be dosed without immediately causing problems.

That is exactly the sort of problem machine learning should be good at. Toxicity, ADMET, binding behaviour, structure-activity relationships - these are data-rich, pattern-heavy domains. Feed the models enough examples and they get better at avoiding the obvious landmines.

That matters. It saves time. It saves money. It probably spares some patients from being exposed to bad compounds.

But it does not prove that AI is solving drug development.

It proves that AI is getting better at helping molecules survive the easiest part of the clinical funnel.

Phase II is where the story gets slippery

Phase II is where the narrative becomes much less clean.

Some analyses suggest AI-backed candidates are succeeding in Phase II at 65 to 75 percent, compared with something like 30 to 45 percent historically. If that holds, it would be a very big deal.

Other, more cautious analyses - usually working from smaller samples - put AI-discovered Phase II success nearer 40 percent, which is basically indistinguishable from the historical range.

This is not just statistical noise. It is a taxonomy problem.

“AI-discovered” currently means almost anything anyone wants it to mean.

It can mean a molecule generated de novo by a model. It can mean a conventional candidate that was later optimised with AI. It can mean AI helped pick the target, refine the chemistry, screen the compound, predict toxicity, or nudge a pre-existing programme over the line. Those are not the same claim.

If a human team chooses the target, identifies the biological rationale, selects the candidate, and then uses AI to optimise some chemistry, is that an AI-discovered drug? Maybe. But it is not the same thing as a model discovering a novel therapeutic hypothesis and producing a molecule that translates into clinical benefit.

Pool all of that together and you can manufacture almost any headline you want.

“AI drugs are doubling Phase II success.”

“AI drugs are no better than traditional drugs.”

Both can be directionally defensible, depending on what gets counted.

And then there is the one number that should keep everyone honest: zero.

As of today, no AI-designed drug has been approved by regulators.

The strongest efficacy signal so far is Insilico Medicine’s rentosertib in idiopathic pulmonary fibrosis, which produced positive Phase IIa results published in Nature Medicine. That is meaningful. It may be the first real proof point that AI-designed molecules can do more than look clean in early safety studies.

But one data point is not a platform shift.

It could be a flare in the dark.

Phase I was never the scarce resource

Here is the question I think the industry is avoiding:

What if AI drug discovery is working - but working on the wrong problem?

Getting molecules into Phase I has never been pharma’s scarce resource.

Large pharmaceutical companies have not historically been sitting around saying: “If only we had more plausible compounds to test in humans.” They have had plenty of plausible compounds. What they have not had is infinite capital, infinite trial capacity, infinite clinical operations bandwidth, or infinite patients willing and eligible to enrol.

The scarce resource is not the mouth of the funnel.

It is the neck.

Drug development gets truly expensive when you have to show efficacy in messy human beings over time. Not binding in a model. Not target engagement in a small trial. Not tolerability in healthy volunteers. Actual clinical benefit, in heterogeneous patients, across sites, geographies, comorbidities, background therapies, adherence problems and all the other inconvenient realities that make biology so rude.

That is where the bodies are buried.

Phase II and Phase III are where capital gets incinerated. They are where management conviction meets patient heterogeneity. They are where elegant biology runs into disappointing effect sizes. They are where the industry discovers that a mechanism can be true without being useful.

This is why the “more shots on goal” metaphor has always bothered me.

In football, more shots are good because the pitch, the goal and the keeper stay roughly the same size.

In pharma, the goal moves. The keeper changes. Half the shots are taken in fog. And by the time you find out whether the shot was any good, you may have spent $500 million and seven years.

The bottleneck was never the number of shots.

It was the conversion rate.

The optionality defence

There is, however, a more charitable version of the AI drug discovery story.

And I think it is worth taking seriously.

Early-stage drug development is not only a funnel. It is also an options portfolio.

A preclinical asset, or even a Phase I candidate, is not a commitment to spend hundreds of millions of dollars. It is a call option on a future medicine. You pay a relatively small premium for the right to learn more. If the signal strengthens, you keep going. If the signal weakens, you stop.

In that framing, AI does not need to produce approved drugs immediately to be useful.

It can still create value by making options cheaper, faster and more varied.

That is the strongest version of the bull case.

AI can generate more hypotheses. It can explore more chemical space. It can produce cleaner molecules earlier. It can identify dead ends before humans would have found them. It can create more shots - but more importantly, it can make each shot less expensive to take.

That is real.

But optionality only works if you preserve the right not to exercise the option, which is where the story gets dangerous.

The value of a cheap option is not that it gives you an excuse to keep investing. The value is that it lets you learn disproportionately more than you spend. It increases the surface area for discovery without forcing you to drag every asset into the expensive part of the funnel.

AI is useful if it makes early-stage assets cheaper to create and cheaper to kill. It is dangerous if it only makes them cheaper to create. Because then optionality becomes inventory.

Inventory in pharma is not harmless. Every programme that survives too long competes for management attention, translational resources, clinical operations bandwidth, trial sites, patients and capital. A large early pipeline can look like strategic abundance while quietly becoming downstream congestion.

The question, then, is not whether AI gives pharma more options.

It probably does.

The question is whether it improves the discipline with which those options are exercised.

A wider funnel can make the bottleneck worse

This is where the AI story gets more interesting - and more uncomfortable.

If AI helps generate and advance more candidates into the clinic, while Phase II and Phase III capacity remains constrained, then the system has not been fixed. It has been loaded.

More molecules arrive at the expensive part of the process. More programmes compete for the same patients, the same investigators, the same trial sites, the same CRO capacity, the same capital committees and the same regulatory bandwidth.

A wider funnel mouth does not help if the neck has not changed shape.

It just creates a queue.

And queues create bad incentives.

Portfolio managers want momentum. Boards want pipeline depth. Investors want clinical-stage assets. Founders want to show that the model works. Everyone has a reason to push one more candidate forward, because the marginal molecule now looks cheaper, cleaner and more defensible than it used to.

But if the downstream system is capacity-constrained, the result is not necessarily better drug development.

It may be more underpowered trials, faster go/no-go decisions, noisier readouts, thinner resourcing per asset, and more capital spread across programmes that should probably have been killed earlier.

This is the paradox. AI may reduce early attrition while increasing downstream congestion. That is not a failure of the technology. It is a failure to understand where the constraint actually sits.

The real test starts now

None of this means AI drug discovery is fake. That would be the lazy sceptical take that I’ve seen on X a lot, and I do not think it is right.

The Phase I data matters. Better safety prediction matters. Faster chemistry matters. Cheaper iteration matters. A world where fewer toxic or badly behaved molecules reach humans is unambiguously better than the alternative.

The question is whether those improvements compound into better medicines, or simply create a more efficient preclinical-to-Phase-I machine.

That is why 2026 is such an important year. The 15 to 20 AI-discovered programmes expected to move into Phase III are the real experiment. Not the press releases. Not the Phase I survival curves. Not the platform demos. Phase III.

This is where small early signals are forced to survive contact with large, diverse populations - - to create more competitive labels. It is where follow-up gets longer, endpoints get harder, adherence gets messier, and efficacy has to be real enough to matter.

If AI-discovered drugs start clearing Phase III at meaningfully better rates than history, then the story changes.

That would suggest AI is not merely filtering for cleaner molecules. It would suggest it is improving the upstream choices that actually matter - target selection, patient segmentation, mechanism validation, biomarker strategy, and the match between molecule, disease and trial design, which would be a very big deal.

But if these programmes fail at roughly normal rates, the conclusion will be just as important. It would mean AI has made the front end of the funnel more efficient without materially changing the part of the system that decides whether a drug reaches patients.

That would still be useful. It just would not be the revolution everyone has been selling.

What I am watching

The next two years should tell us much more than the last ten.

First, I would watch the Phase III cohort. Not whether a few AI-discovered drugs enter pivotal trials - that bar is now being cleared. The question is whether they succeed at rates that are meaningfully different from comparable historical programmes.

Second, I would watch whether the industry gets serious about definitions. “AI-discovered” is currently doing far too much work. A molecule designed de novo by a model is not the same thing as a conventional asset polished with AI chemistry tools. Until the field has a credible taxonomy, every aggregate success-rate statistic will be a marketing asset first and an analytical object second.

Third, I would watch kill rates and decision quality before Phase II and Phase III. If AI is creating valuable optionality, we should not only see more candidates entering the clinic. We should see better-ranked portfolios, faster termination of weak programmes, clearer biomarker strategies, and more disciplined decisions about which assets deserve scarce late-stage capacity. The positive signal is not a (more) swollen pipeline. It is a pipeline that learns faster.

Fourth, I would watch capacity strain. If AI keeps increasing the number of Phase I-cleared candidates without a matching increase in patient recruitment, trial-site capacity and late-stage funding discipline, the symptoms should become visible: thinner trials, stretched CROs, faster portfolio churn, and more ambiguous readouts.

Fifth, I would watch whether rentosertib gets a sibling. One positive Phase IIa result in Nature Medicine is important. A second independent efficacy signal, from a different company in a different disease area, would be much more important. That is when anecdotes start becoming a pattern.

The bottom line

AI is making pharma better at producing plausible clinical candidates. That is real progress. But plausible clinical candidates were not the thing pharma was shortest of.

The industry’s hardest problem is not designing molecules that look good enough to enter humans. It is proving that those molecules change outcomes in real patients, at scale, over time, strongly enough to justify approval, reimbursement and use.

The optimistic case is that AI turns the early funnel into a cheaper learning machine - more hypotheses tested, more dead ends killed early, more capital reserved for the few assets that genuinely deserve it. The pessimistic case is that it turns optionality into inventory.

A wider funnel mouth is only useful if it produces better decisions at the neck.

So the right question after every “AI drug enters Phase I” headline is not: did AI discover another molecule? It is: What did we learn, how cheaply did we learn it, and did it change what we chose not to do next?

At one level, that is true. A molecule is not a software roadmap. Management cannot will a biological outcome into existence. The mechanism will translate, or it will not. The safety profile will emerge, or it will not. The data will say what the data says.

But that objection misses where optionality actually sits in biotech.

Optionality is not about controlling the biology. It is about controlling the learning agenda around the biology. And in practice that matters because earlier, deeper learning can materially increase the pre-launch value of an asset.

It does so in two ways. First, it increases confidence in more than one strategically relevant direction. Secondly, it forces harder questions earlier, when the cost of answering them - and acting on the answers - is still relatively low.

That is not a philosophical distinction. It has direct implications for programme design, financing strategy, business development, and how an asset is priced by the market.

Biology is fixed. Strategic range is not.

The phrase “the drug will do what it will do” collapses two different issues into one.

The first is scientific uncertainty: what does the molecule actually do, in which patients, under which conditions, and with what magnitude of effect?

The second is strategic range: given what the biology reveals, how many credible paths does the company have to create value?

That range may include:

• a faster first indication;

• a narrower but more actionable biomarker-defined population;

• a better sequencing strategy;

• a combination path;

• a partnering route with clearer proof points;

• or, just as importantly, an earlier decision to stop.

The molecule may be fixed. But the set of credible actions available to the company is not fixed at the outset. It is shaped by what the company chooses to learn, when it chooses to learn it, and how well that learning is translated into a development and partnering strategy.

That is the real source of optionality.

The real risk is not lack of focus - it is false precision

Small biotechs are often told to stay focused. That advice is not wrong, but it is frequently misapplied.

In practice, “focus” can mean one of two very different things.

The good version is disciplined early design: asking the most decision-relevant questions as efficiently as possible.

The bad version is premature narrowing: building the company around one simplified proof-of-concept story before the underlying asset has been characterised deeply enough to justify that level of certainty.

That second version often looks tidy in a pitch deck. It can also be value-destructive.

A narrow early plan may produce a clean data readout for the next round. But it can leave critical questions unanswered:

• Is the lead indication actually the best entry point?

• Is the effect concentrated in a subgroup the current design will blur?

• Is there a biomarker strategy that would materially improve the asset’s positioning?

• Are there adjacent indications with faster regulatory or commercial paths?

• Are there mechanism or safety signals that will matter later, but are cheaper to investigate now?

When those questions are postponed, they do not disappear. They simply become more expensive - scientifically, financially, and strategically.

How deeper early exploration creates asset value

The case for richer early exploration is not that more data is always better. It is that well-chosen early information increases the number and quality of credible future moves.

There are two main ways it does this.

1. It creates confidence in more than one value-creating path

An early programme that includes the right biomarker work, translational measures, secondary endpoints, or adaptive cohorts can do more than validate a single headline thesis.

It can also clarify:

• where the biology is strongest;

• which patients are most likely to respond;

• whether there is a narrower, faster, or more defensible first indication;

• whether the mechanism has relevance beyond the initial target setting;

• and whether the asset supports a broader platform or follow-on story.

This is not “spray and pray” exploration. It is targeted effort designed to preserve credible choices until the evidence justifies narrowing them.

That matters because asset value is rarely driven by a single binary question alone. It is also driven by what a credible counterparty - investor, partner, or acquirer - believes could happen next.

A molecule with one plausible path may be interesting. A molecule with one plausible path plus two credible expansions, a clearer responder logic, and a better-defined development sequence is usually more valuable before approval, even if the underlying biological signal is unchanged.

2. It surfaces hard questions while they are still cheap

The second benefit is often underappreciated.

Deeper early work does not just reveal upside. It also reveals constraints.

It forces management, boards, and investors to confront questions that otherwise get deferred:

• Does the mechanism really translate in the clinically relevant population?

• Is the therapeutic window wide enough for the intended setting?

• Does the biology point towards monotherapy, combinations, or a defined subgroup?

• Are there resistance or durability issues that change the commercial case?

• Is the large headline indication actually less attractive than a smaller, faster, better-defined one?

These are not side questions. They are the questions that determine whether a programme compounds value or burns capital.

When surfaced early, they improve asset allocation. Weak assets can be stopped sooner. Better indications can be prioritised earlier. Trial designs can be improved before large commitments are made. And partnering discussions can be framed around evidence rather than aspiration.

In other words, early depth does not just create upside optionality. It also reduces the cost of being wrong.

Why founders, BD teams, and investors should care

This matters differently depending on where you sit, but the underlying economic logic is the same.

For founders and management teams

Deeper early exploration improves strategic choice.

It can help a company avoid overcommitting to the wrong lead indication, identify a faster path to clinically meaningful proof, and build a more resilient narrative for future financing. It also reduces the odds that the company reaches a major inflection point with a dataset that is clean but strategically thin.

The practical benefit is not complexity for its own sake. It is better decision-making under capital constraints.

For BD teams

A richer early package makes an asset easier to position and easier to transact.

Partners do not just assess whether a programme has positive data. They assess whether they understand the mechanism, the responder logic, the expansion potential, the development risks, and the work still required after signing.

The more of that is clarified early, the more credible the asset becomes in a competitive process. That can improve terms, widen the pool of interested counterparties, and shift negotiations from “help us figure out what this is” to “here is where this can go next”.

For investors

The value of an early asset is not just the probability that the first proof-of-concept readout works. It is the probability-weighted value of the future strategic tree that the asset supports.

That includes upside breadth, downside containment, financing efficiency, partnering leverage, and the quality of future decision points.

An asset that has been explored narrowly may still produce a positive signal. But if it leaves too many high-value questions unanswered, the investor is still underwriting substantial hidden uncertainty. By contrast, a company that has used early capital to improve the shape of the uncertainty may deserve a better valuation even before the ultimate clinical answer is known.

This is not a call for undisciplined expansion

None of this means early programmes should become sprawling.

It does not mean adding endpoints indiscriminately, chasing every weak signal, or pretending every molecule is a platform. And it certainly does not mean small biotechs should spend as if capital were abundant.

The point is more practical than that.

The goal is to identify the highest-leverage early questions - the ones that can most improve asset value, development quality, or negotiating power per pound spent - and design the programme so those questions are answered as early as reasonably possible.

Sometimes that means biomarker work. Sometimes it means richer translational sampling, smarter cohorts, better endpoint architecture, or cleaner archival of data for later repositioning. The specific tools matter less than the discipline behind them.

The best early programmes buy information asymmetrically

That, ultimately, is the point.

The best early biotech programmes do not just buy progress. They buy information asymmetrically.

They spend relatively modest amounts of capital to answer questions that would become far more expensive later. They improve the company’s ability to choose between indications, refine development strategy, structure partnerships, and abandon weak paths before they consume disproportionate resources.

So when someone says a biotech cannot afford to think in options, I think the more relevant question is: can it afford not to create them where they are cheapest?

Not because optionality changes the biology. It does not.

But because it changes how much strategic value a company can extract from what the biology eventually reveals.